Paper deep dive
MINDSET: Energy-based Schema Evolution for Long Conversational Agent Memory
Sujato Dutta, Sreekruthy Tummala, Shashank Vanga, Ayushmi Pavani
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Long conversational agents have become essential in our daily lives. They must remember what was said long back in order to help us efficiently complete a task without needing the user to repeat instructions and context repeatedly. However, the main issue is that instructions and context change over time and so the agents must be able to adapt accordingly. A useful memory system should preserve both current and historical states, distinguish stale information from active knowledge, retrieve evidence appropriate to the query and avoid repeatedly invoking a large language model to rewrite prior interactions. We introduce MINDSET, a memory controller that stores a conversation as immutable episodes and organizes them into versioned schemas through minimum-energy state transitions. Each incoming episode may reinforce, supersede, split or create a schema. The transition decision balances representation distortion, contradiction, historical damage, fragmentation and internal inconsistency, while hysteresis prevents isolated contradictions from prematurely rewriting stable memory. We evaluate MINDSET against 5 memory systems on a reproducible sample of 850 questions (700 LoCoMo + 150 MemoryAgentBench). MINDSET obtains the highest observed LoCoMo answer F1 while significantly improving retrieval ranking (Recall@8, MRR and nDCG@8) over the second best method LightMem (p<0.01 after Holm correction). It obtains the highest observed scores on MemoryAgentBench although the relative difference is low. Ablations identify controlled fragmentation and schema-aware assignment as the largest contributors to answer quality. Additionally, a 700-question cross-model evaluation with GLM-4.7 and Gemma-4-31B supported model independence. These results show that long-term memory can be better handled as constrained state management rather than continual summarization.
Tags
Links
- Source: https://arxiv.org/abs/2610.08586v1
- Canonical: https://arxiv.org/abs/2610.08586v1
Trouble viewing inline? Open PDF directly →
Full Text
56,754 characters extracted from source content.
Expand or collapse full text
MINDSET: Energy-based Schema Evolution for Long Conversational Agent Memory Sujato Dutta Sreekruthy Tummala Shashank Vanga & Ayushmi Pavani Affiliation: Mahindra University Abstract Long conversational agents have become essential in our daily lives. They must remember what was said long back in order to help us efficiently complete a task without needing the user to repeat instructions and context repeatedly. However, the main issue is that instructions and context change over time and so the agents must be able to adapt accordingly. A useful memory system should preserve both current and historical states, distinguish stale information from active knowledge, retrieve evidence appropriate to the query and avoid repeatedly invoking a large language model to rewrite prior interactions. We introduce MINDSET, a memory controller that stores a conversation as immutable episodes and organizes them into versioned schemas through minimum-energy state transitions. Each incoming episode may reinforce, supersede, split or create a schema. The transition decision balances representation distortion, contradiction, historical damage, fragmentation and internal inconsistency, while hysteresis prevents isolated contradictions from prematurely rewriting stable memory. We evaluate MINDSET against 5 memory systems on a reproducible sample of 850 questions (700 LoCoMo + 150 MemoryAgentBench). MINDSET obtains the highest observed LoCoMo answer F1 while significantly improving retrieval ranking (Recall@8, MRR and nDCG@8) over the second best method LightMem (p<0.01 after Holm correction). It obtains the highest observed scores on MemoryAgentBench although the relative difference is low. Ablations identify controlled fragmentation and schema-aware assignment as the largest contributors to answer quality. Additionally, a 700-question cross-model evaluation with GLM-4.7 and Gemma-4-31B supported model independence. These results show that long-term memory can be better handled as constrained state management rather than continual summarization. 1 Introduction A conversational agent must remember not only what was said, but also what changed over time. With changing preferences, retaining only the original preference will produce stale answers, while retaining only the latest statement erases the context required to answer questions about the transition. Storing both in an undifferentiated vector index leaves the model to resolve the contradiction repeatedly. Full-history prompting preserves the complete record, but incurs recurring inference cost and remains vulnerable to context limits and the “lost in the middle” effect Liu et al. (2024). Summarization reduces context length but may blur the context needed to verify an answer. This problem is common in agent memory. Retrieval-augmented generation (RAG) Lewis et al. (2021) provides language models with external evidence, while dense passage retrieval Karpukhin et al. (2020) and sentence embeddings Reimers and Gurevych (2019) make semantic access efficient. Memory-based agents have introduced reflection Park et al. (2023); Shinn et al. (2023), virtual context management Packer et al. (2024), forgetting Zhong et al. (2023), graph-based organization Gutiérrez et al. (2025), learned memory operations Yu et al. (2026) and benchmark-specific consolidation Zhang et al. (2026); Ma et al. (2026). These systems show the importance of memory but lack in a core area. The central challenge in long conversational agents is state evolution under contradiction i.e. deciding whether new evidence should update an existing one, coexist as a contextual branch or be stored as a new memory. MINDSET addresses this problem directly. It maintains immutable conversational episodes and a compact layer of versioned schemas. A schema is not a generated summary; it is a structured index over supporting and exceptional episodes, with temporal validity, active or historical status and predecessor-successor links. For each incoming episode, MINDSET evaluates four possible actions - reinforce, supersede, split and new under a local energy function. The energy penalizes representation distortion, unresolved contradiction, historical damage, unnecessary fragmentation, and internal inconsistency. Temporal and contextual signals determine action feasibility, while hysteresis prevents isolated contradictions from prematurely rewriting stable memory. This design separates two responsibilities that are often conflated. The memory controller determines how experience should be organized and which episodes are relevant, while the language model receives a bounded evidence packet and produces the final answer. Our main contributions are: 1. We formulate conversational memory evolution as minimum-energy selection over explicit state-transition actions. 2. We introduce provenance-preserving versioned schemas that distinguish active states, historical states, contextual branches and exceptions without replacing the underlying episodes. 3. We provide a thorough and fair evaluation across LoCoMo and MemoryAgentBench, covering six memory systems, ablations and 3 answer models to show generalization. Our results show that MINDSET achieves the best retrieval-ranking performance along with best overall answer F1. It also lowers token usage by upto 10 times when compared with long context. 2 Related Work 2.1 Retrieval and Long-Context Answering Retrieval-augmented generation (RAG) combines parametric generation with non-parametric evidence Lewis et al. (2021). Dense passage retrieval Karpukhin et al. (2020) established a widely used neural retrieval paradigm, Sentence-BERT Reimers and Gurevych (2019) made efficient semantic matching practical, and BM25 Robertson and Zaragoza (2009) remains competitive when exact entities, dates, or rare lexical cues are important. Reciprocal rank fusion Cormack et al. (2009) combines heterogeneous rankings without requiring score calibration, while maximal marginal relevance Carbonell and Stewart (1999) balances relevance and diversity. MINDSET builds on this and modifies the memory index: raw episodes are supplemented with versioned schema routes, historical-state access and provenance expansion back to the original evidence. Long-context prompting provides an alternative to explicit memory management. However, evaluations such as LongBench Bai et al. (2024), RULER Hsieh et al. (2024), and Lost in the Middle Liu et al. (2024) show that larger context windows do not guarantee reliable evidence use. Long inputs also incur recurring inference cost. 2.2 Memory for Language-Model Agents Early agent-memory systems demonstrated several complementary strategies. Generative Agents stores observations, generates reflections, and retrieves memories to support behavior Park et al. (2023); ReAct interleaves reasoning with environment-facing actions Yao et al. (2023); MemGPT treats context management as a virtual-memory problem Packer et al. (2024); and Reflexion stores linguistic feedback as episodic memory rather than modifying model weights Shinn et al. (2023). MemoryBank combines dialogue memories, summaries and personality descriptions Zhong et al. (2023). More recent work explores increasingly structured forms of memory control. A-MEM dynamically organizes agent-generated notes Xu et al. (2025), MemoryOS introduces an operating-system abstraction Kang et al. (2025), Mem0 emphasizes scalable extraction and consolidation Chhikara et al. (2025), HippoRAG uses graph-based associative retrieval Gutiérrez et al. (2025), and AgeMem learns unified memory operations Yu et al. (2026). Among systems closest to our experimental setting, LightMem separates short, mid and long-term memory while emphasizing lightweight memory operations Zhang et al. (2026), whereas Nemori organizes experience through episodic boundaries, aligned representations, and predict-calibrate updates Ma et al. (2026). 2.3 Evolving and Temporal Memory Temporal change is central to several recent benchmarks. LoCoMo evaluates single-hop, temporal, multi-hop, open-domain, and adversarial reasoning over long conversations Maharana et al. (2024). LongMemEval studies retention, multi-session reasoning, temporal reasoning, knowledge updates, and abstention across extended interaction histories Wu et al. (2025). Beyond Goldfish Memory introduced long-running multi-session dialogue Xu et al. (2022), LOCCO measures chronological degradation Jia et al. (2025), and MemBench broadens evaluation across effectiveness, efficiency, and capacity Tan et al. (2025a). Keep Me Updated targets changed world knowledge Bae et al. (2022), while systems such as RMM Tan et al. (2025b), THEANINE Ong et al. (2025), Associa Zhang et al. (2025), EMG-RAG Wang et al. (2024), HiAgent Hu et al. (2025), MemoryOS Kang et al. (2025), and MemInsight Salama et al. (2025) explore reflective, hierarchical, associative, graph, and diagnostic views of memory. MemoryAgentBench evaluates factual retrieval, in-context learning, long-range understanding, and recommendation Hu et al. (2026). MINDSET makes memory evolution an explicit choice among four state transitions: reinforcing an existing schema, superseding a historical state, creating a contextual branch and starting a new schema. This formulation is related in spirit to event segmentation in cognition Zacks and others (2007), complementary learning systems Mcclelland et al. (1995), and energy-based learning LeCun et al. (2006). We do not claim a biologically faithful model. Instead, they motivate the engineering principle behind MINDSET. 3 Method Figure 1: MINDSET stores immutable episodes, chooses minimum-energy transitions over versioned schemas, expands retrieved schemas back to provenance, and sends a bounded raw-evidence packet to the answer model. 3.1 Problem formulation Let a conversation produce an ordered stream of episodes ℰt=(e1,e2,…,et),E_t=(e_1,e_2,…,e_t), where each episode stores a conversation identifier, session, timestamp, speaker, raw text and locally extracted entity, role, context, attribute, and change-cue features. The memory controller maintains a set of schemas tS_t. A schema groups compatible episodes into a versioned state; it records support and exception identifiers, a centroid and medoid, validity bounds, active or historical status and predecessor/successor links. For each incoming episode ete_t, the controller chooses an action at∈reinforce,supersede,split,newa_t∈\reinforce,supersede,split,new\ Reinforce adds evidence to a compatible active schema. Supersede closes an active schema and creates its temporally linked successor. Split creates a coexisting contextual branch. New starts an independent schema. Raw episodes are immutable under every action. 3.2 Compatibility and transition signals MINDSET first retrieves at most ksk_s nearby active schemas. Let x be the unit embedding of incoming episode e, s μ_s the unit centroid of schema s, and nsn_s its number of supporting episodes. Dense similarity is d=max(0,cos(,s))d= (0, (x, μ_s)). For a schema feature counter ng(v)n_g(v) and the incoming set Xg(e)X_g(e), the implementation uses Og(s,e)=o∅,∑vng(v)=0 or Xg(e)=∅,∑v∈uniq(Xg(e))ng(v)∑vng(v),otherwise.O_g(s,e)= \ array[]lo_ , _vn_g(v)=0 or X_g(e)= ,\\[8.53581pt] _v (X_g(e))n_g(v) _vn_g(v), . array . Entity, role, and context matches are oe=Oentityo_e=O_ entity, or=Oroleo_r=O_ role, and mc=Ocontextm_c=O_ context. The topic indicator mtm_t is one when the episode’s primary extracted attribute slot is not discourse and equals the schema topic suffix, and zero otherwise. Compatibility is exactly K(e,s)=clip[0,1](CLOSE K(e,s)=clip_[0,1] ( αdd+αeoe+αror _dd+ _eo_e+ _ro_r OPEN+αcmc+αtmt). + _cm_c+ _tm_t ). Contradiction is computed only for incoming single-valued slots already present in the schema. Let Vs(q)V_s(q) and Ve(q)V_e(q) be the existing and incoming values for slot q, and let J(u,v)J(u,v) be Jaccard similarity between their normalized non-stopword token sets (exact string equality is used when either set is empty). With preference, plan, possession, experience, ability, and dislike treated as multi-valued and excluded, p=maxq[1−maxu∈Vs(q),v∈Ve(q)J(u,v)],p= _q [1- _u∈ V_s(q),v∈ V_e(q)J(u,v) ], with p=0p=0 when no eligible slot exists. Recurrence is r=min(1,Mmr),r= (1, Mm_r ), where M is the number of stored exception episodes containing a same-slot value with J≥θrJ≥ _r. Let ℓ indicate that the incoming timestamp is no earlier than the latest support timestamp, and let x indicate an explicit lexical change cue. The implemented temporal and contextual potentials are τ= τ= pmin1,βcmc+βℓ p \1,\, _cm_c+ _ +βx+βr, + _xx+ _rr \, κ= κ= max(p,γtmt) (p, _tm_t) ×min1,γc(1−mc) × \1,\, _c(1-m_c) +γr(1−or)+γd(1−d). + _r(1-o_r)+ _d(1-d) \. These signals determine action feasibility before energy comparison: supersede requires p≥θsupp≥ _ sup, split requires κ≥θsplitκ≥ _ split, and new is removed when K≥θnewK≥ _ new. A hard new-topic guard selects new when K<θnewK< _ new, mt=0m_t=0, and d<θd< _d. The implementation uses ks=6k_s=6, o∅=0.50o_ =0.50, (αd,αe,αr,αc,αt)=(0.52,0.18,0.12,0.10,0.08)( _d, _e, _r, _c, _t)=(0.52,0.18,0.12,0.10,0.08), mr=2m_r=2, θr=0.80 _r=0.80, (βc,βℓ,βx,βr)=(0.35,0.20,0.40,0.25)( _c, _ , _x, _r)=(0.35,0.20,0.40,0.25), (γt,γc,γr,γd)=(0.65,0.45,0.20,0.35)( _t, _c, _r, _d)=(0.65,0.45,0.20,0.35), and (θsup,θsplit,θnew,θd)=(0.35,0.12,0.43,0.35)( _ sup, _ split, _ new, _d)=(0.35,0.12,0.43,0.35). All compatibility coefficients, transition thresholds, energy weights and hysteresis margins were selected through coarse calibration on a held-out development set with no question overlap with LoCoMo-700. MAB is fully dataset-disjoint from this calibration. They were frozen before test execution and were not optimized against test-answer F1. Episode Conversation excerpt K p τ κ Relevant feasible energies Decision D3:16 “Setting small goals and tracking my progress…” .535 .000 .000 .000 reinforce .587 reinforce D3:18 “I’m getting into different types of games now…” .622 1.000 .775 .378 reinforce 2.482; supersede .094; split .534 supersede D14:2 “Last week, I started my blog about coding.” .418 .000 .000 .000 reinforce .631; new .549 new D18:1 “Recently left my IT job…now with this new job…” .480 .000 .000 .351 reinforce .625; split .181 split Table 1: Audited transitions from one frozen LoCoMo conversation. Energies are shown only for feasible actions. 3.3 Minimum-energy schema evolution For every feasible action, MINDSET computes E(a)= E(a)= wdDa+wcCa+whHa w_dD_a+w_cC_a+w_hH_a +wfFa+wiIa. +w_fF_a+w_iI_a. Every term below is the quantity computed by the released implementation. Define + μ^+ =unit(nss+), =unit (n_s μ_s+x ), δs _s =1−cos(s,+), =1- ( μ_s, μ^+ ), de+ d_e^+ =1−cos(,+). =1- (x, μ^+ ). If D¯s D_s is the schema’s running mean distortion, Da= D_a= [a=reinforce] [a=reinforce] ×ns(D¯s+δs)+de+ns+1. × n_s( D_s+ _s)+d_e^+n_s+1. Unresolved contradiction and historical damage are Ca C_a =[a=reinforce]p, =I[a=reinforce]p, Ha H_a =[a=reinforce](δs+λhp). =I[a=reinforce]( _s+ _hp). Contradiction is therefore charged when incompatible evidence is merged; superseding or branching preserves the old state rather than paying this merge cost. For fragmentation, let A be the active-schema count, T the count with fewer than mminm_ support episodes, N the registered episode count including e at transition evaluation, qa q_a =[a∈split,new], =I\! [a∈\split,new\ ], ba b_a =[a=reinforce][ns<mmin≤ns+1]. =I[a=reinforce]\,I[n_s<m_ ≤ n_s+1]. Then Fa=A+qa+η(T+qa−ba)max(1,N+1).F_a= A+q_a+η(T+q_a-b_a) (1, N+1). Here η is a coefficient on projected tiny schemas, not an exponent. Let sG_s contain every nonempty attribute, role, and context counter in s, and define U(s)= U(s)= 1|s|∑g∈s(1−maxvng(v)∑vng(v)), 1|G_s| _g _s (1- _vn_g(v) _vn_g(v) ), with U(s)=0U(s)=0 when sG_s is empty. The action-specific inconsistency term is Ia=U(s)+λimax(τ,κ),a=reinforce,max(0,κ−τ),a=supersede,max(0,τ−κ),a=split,K(e,s),a=new.I_a= \ array[]lU(s)+ _i (τ,κ),&a=reinforce,\\[2.84526pt] (0,κ-τ),&a=supersede,\\[2.84526pt] (0,τ-κ),&a=split,\\[2.84526pt] K(e,s),&a=new. array . The final weights are represented by =(wd,wc,wh,wf,wi).w=(w_d,w_c,w_h,w_f,w_i). The implementation uses λh=0.35 _h=0.35, mmin=2m_ =2, η=0.75η=0.75, λi=0.50 _i=0.50, and =(1.00,1.25,0.75,0.22,1.00).w=(1.00,1.25,0.75,0.22,1.00). These are hand-tuned coefficients, not a proven optimal fit. Distortion and inconsistency define unit scale; contradiction receives modest priority so a detected single-valued conflict can outweigh moderate semantic fit; historical damage is discounted; and fragmentation is a weaker global regularizer because its count-based term accumulates throughout a conversation. The development set is question disjoint from test LoCoMo 700. There is no gradient fitting or test-answer F1 optimization. All parameters are frozen before the final paper runs. Two shortcuts precede general comparison. Easy reinforcement applies when p<θpeasyp< _p easy, κ<θκeasyκ< _κ easy, and K≥θKeasyK≥ _K easy. An isolated strong contradiction (p≥θpiso,r=0,x=0p≥ _p iso,r=0,x=0) is also assigned to reinforce. Whenever the final action is reinforce and p≥θexcp≥ _ exc, the episode is stored as an exception rather than support. Otherwise let b be the lowest-energy feasible alternative and Δ=E(reinforce)−E(b). =E(reinforce)-E(b). The effective switch margin is meff= m_ eff= mswitch(1−λrr) m_ switch(1- _rr) ×ρ,τ>κ and τ>θτ,1,otherwise. × \ array[]lρ,&τ>κ and τ> _τ,\\[2.84526pt] 1,&otherwise. array . We use (θpeasy,θκeasy,θKeasy)=(0.15,0.12,0.58)( _p easy, _κ easy, _K easy)=(0.15,0.12,0.58), θpiso=0.60 _p iso=0.60, θexc=0.45 _ exc=0.45, mswitch=0.18m_ switch=0.18, λr=0.60 _r=0.60, ρ=0.55ρ=0.55, θτ=0.45 _τ=0.45 and mstay=0.06m_ stay=0.06. These thresholds and margins were included in the same hand-tuned development protocol. The controller selects b when Δ>meff >m_ eff, retains reinforce when Δ≤mstay ≤ m_ stay, and selects b in the intermediate region. Thus the stay margin is the final action boundary; the switch margin records whether a transition clears the stronger evidence threshold but does not change the action once the stay margin is crossed. 3.4 An audited transition sequence Table 1 follows four real episodes from the frozen LoCoMo conversation conv-47. At D3:16, only reinforce remains feasible. At D3:18, “now” marks an explicit change and a single-valued conflict; supersede costs 0.094, compared with 2.482 for reinforce. D14:2 introduces a coding blog below the 0.43 new-schema threshold, and new beats reinforce by 0.083. D18:1 has contextual potential 0.351 but no eligible attribute contradiction, so split preserves the earlier state while adding a branch. Feasibility and hysteresis are part of the policy: the lowest-energy action among all four alternatives is not necessarily feasible. 3.5 Retrieval from versions without summarizing away evidence At query time, a local router labels the information need as current, historical, contextual or generic. Three candidate channels are then combined: • direct retrieval, a dense-BM25 union over raw episodes; • schema-mediated retrieval, which finds relevant schemas and expands them back to their support, exception, medoid and continuity episodes; • version retrieval, which exposes medoids from active and historical schema states. The candidate pool is reranked with ms-marco-MiniLM-L6-v2 and embeddings use all-MiniLM-L6-v2. We use a candidate pool of 40 and return eight pieces of evidence. The answer model receives the original provenance-bearing episodes. Method Answer F1 Recall@8 Precision@8 MRR nDCG@8 Evidence recall Total tokens MINDSET 0.4280 0.6280 0.1056 0.5572 0.5453 0.6302 403,978 LightMem 0.4024 0.4522 0.0783 0.4365 0.4092 0.4554 313,522 Vector RAG 0.3807 0.3904 0.0670 0.2675 0.2779 0.3939 356,368 Nemori 0.3733 0.4198 0.0682 0.2527 0.2729 0.6321 1,536,928 MemoryBank 0.2975 0.0634 0.0106 0.0438 0.0463 0.6104 1,724,495 Long Context 0.2791 0.0153 0.0036 0.0082 0.0082 0.2190 4,310,788 Table 2: Main LoCoMo-700 results with GPT-OSS-120B. Total tokens are provider-recorded input plus output tokens. Exact cache hits add zero provider tokens in the run ledger; MemoryBank includes both memory-construction and answer-generation tokens. 3.6 Common evidence budget Retrieval methods can emit memory units of very different size. We therefore normalize them before generation. Up to 20 candidate memories are segmented into 768-token passages with 96-token overlap, rescored against the question using the same local embedding model and greedily selected under a maximum of eight passages - 8,000 evidence tokens and three passages per original memory. Every passage retains its source memory ID, source event IDs, passage index, and character/token offsets. Long Context is the sole structural exception because full history defines the baseline; it uses deterministic recency truncation at 20,000 UTF-8 bytes. 3.7 Computational efficiency MINDSET performs no gradient training and makes no provider call during ingestion or retrieval. Its learned components are frozen off-the-shelf local encoders. If n episodes have already been stored, insertion compares a bounded set of nearby active schemas; retrieval constructs bounded direct and schema pools before reranking. The implementation therefore avoids both an unbounded prompt at answer time and repeated generative rewriting at write time. 4 Experimental Setup 4.1 Datasets We use a frozen manifest of 700 questions from LoCoMo Maharana et al. (2024), 70 from each of ten conversations. The category counts are 151 single-hop, 151 temporal, 96 multi-hop, 151 open-domain, and 151 adversarial questions. As a secondary dataset, we use MemoryAgentBench (MAB) Hu et al. (2026). We use 150 samples from benchmark-instance files: 125 in-context-learning examples and 25 recommendation examples. The five ICL source groups contribute 25 instances each, and the recommender contributes 25. We therefore use 850 questions in total. This is done intentionally to keep API costs manageable while also providing sufficient breadth for fair evaluation. The exact IDs and order are stored in the manifest and will be released with the codebase for reproducibility. Model transfer is evaluated on the complete frozen LoCoMo benchmark used by the primary experiment. 4.2 Baselines We compare MINDSET against 5 methods: • Vector RAG: flat normalized MiniLM embeddings over timestamped dialogue turns with cosine top-k retrieval. • Long Context Memory: chronological full history, with deterministic oldest-turn removal only beyond 20,000 UTF-8 bytes. • MemoryBank: adapted from the official architecture Zhong et al. (2023). GPT-OSS replaces the original GPT-3.5 summarizer so the backbone remains same. • LightMem: a reproduction preserving key design choices from the paper Zhang et al. (2026). • Nemori: a reproduction of the official implementation Ma et al. (2026). 4.3 Answer models and generation The primary model is gpt-oss-120b from Cerebras. All six methods use the same answer instructions, deterministic temperature 0, a maximum output reservation of 4,096 tokens, structured JSON validation, and completeness retries. Primary LoCoMo batching uses eight questions where supported; Long Context uses one question per request because histories are question-specific and large. The cross-backbone experiment uses zai-glm-4.7 and gemma-4-31b from Cerebras, 1 question per request, a 1,024-token output cap and a 160,000-byte prompt guard. Method ICL exact match Recommender Recall@5 ICL token F1 Recommender F1@5 MINDSET (Ours) 0.0240 0.1400 0.0526 0.0587 MemoryBank 0.0000 0.1000 0.0409 0.0427 LightMem 0.0160 0.0400 0.0414 0.0160 Long Context Memory 0.0080 0.0400 0.0275 0.0133 Vector RAG 0.0000 0.0400 0.0183 0.0133 Nemori 0.0240 0.0000 0.0506 0.0000 Table 3: GPT-OSS-120B on the frozen MAB-150 manifest. Variant Answer F1 Recall@8 Precision@8 MRR nDCG@8 Evidence recall Full MINDSET 0.4280 0.6280 0.1056 0.5572 0.5453 0.6302 Nearest-schema merge 0.4080 0.6296 0.1060 0.5547 0.5445 0.6318 No contextual term 0.4194 0.6280 0.1056 0.5572 0.5453 0.6302 No temporal term 0.4237 0.6266 0.1054 0.5556 0.5437 0.6287 No hysteresis 0.4237 0.6280 0.1060 0.5549 0.5442 0.6302 No fragmentation penalty 0.4043 0.6256 0.1052 0.5567 0.5437 0.6278 No provenance expansion 0.4274 0.6158 0.1029 0.5508 0.5369 0.6180 Table 4: LoCoMo-700 ablations with GPT-OSS-120B. 4.4 Metrics For LoCoMo we report answer F1 and retrieval Recall@8, Precision@8, mean reciprocal rank (MRR), nDCG@8 Järvelin and Kekäläinen (2002) and evidence recall. Text normalization removes punctuation and articles and applies Porter stemming. Temporal, multi-hop, and open-domain questions use token F1; single-hop questions admit multiple reference answers; adversarial questions are correct only when the response abstains with "No information available" or equivalent benchmark-recognized wording. MAB uses its native metrics: first-line exact match for ICL and Recall@5 for recommendation. We additionally report deterministic normalized token F1 for ICL and set F1@5 for recommendation. 5 Results In all tables, bold marks the best result and blue underlining marks the second-best. Ties receive the same rank formatting. 5.1 Main results on LoCoMo MINDSET obtains the best observed answer F1 although the gap between LightMem stays close. In ranking metrics, MINDSET stays clear of other methods. Its MRR exceeds the runner-up by 0.1207 and its nDCG@8 by 0.1361. Only in evidence recall, it trails behind Nemori but that evidence is less concentrated at useful ranks and yields lower answer F1. In terms of token consumption, MINDSET uses almost 10x less tokens as compared to long context while trailing behind LightMem and Vector RAG. 5.2 Statistical tests On the ten LoCoMo conversation clusters, MINDSET significantly improves Recall@8, MRR and nDCG@8 over LightMem, the closest method across our study (mean differences +0.1758+0.1758, +0.1207+0.1207, and +0.1361+0.1361, respectively; Holm-corrected p=0.00977p=0.00977 for each). Against Nemori, the gains in Recall@8 (+0.2082+0.2082), MRR (+0.3045+0.3045) and nDCG@8 (+0.2724+0.2724) are also significant (Holm-corrected p=0.00977p=0.00977 for each). Thus, the tests support a significant retrieval-ranking advantage over both strongest baselines. For answer F1, MINDSET obtains the highest observed value. Its gain over long context is significant (mean difference +0.1489+0.1489, 95% cluster-bootstrap CI [+0.1094,+0.1857][+0.1094,+0.1857], Holm-corrected p=0.00977p=0.00977). Gains over Vector RAG, Nemori, and MemoryBank are also significant at p<0.05p<0.05 (Holm-corrected p=0.04297p=0.04297, 0.029300.02930, and 0.015620.01562, respectively). The gain over LightMem remains directionally positive but is not statistically confirmed (+0.0256+0.0256, p=0.24023p=0.24023). We use two-sided exact paired sign-flip permutation tests over all 210=1,0242^10=1,024 conversation-cluster sign assignments and cluster bootstrapping at the conversation level, with Holm correction within each metric family. This design accounts for non-independence among questions drawn from the same history. Method Answer F1 ↑ Recall@8 ↑ Precision@8 ↑ MRR ↑ nDCG@8 ↑ Total Tokens ↓ MINDSET 0.4080 0.4293 0.0745 0.2806 0.2957 507,238 LightMem 0.4005 0.4130 0.0718 0.2744 0.2875 708,450 Nemori 0.3961 0.3085 0.0510 0.1794 0.1937 955,224 Vector RAG 0.3815 0.3904 0.0670 0.2675 0.2779 609,491 Long Context 0.2975 0.0153 0.0036 0.0082 0.0082 4,360,926 MemoryBank 0.2659 0.0671 0.0115 0.0501 0.0494 831,244 Table 5: GLM-4.7 on the frozen LoCoMo-700 manifest. Total tokens is the provider-recorded sum of input and output tokens. Method Answer F1 ↑ Recall@8 ↑ Precision@8 ↑ MRR ↑ nDCG@8 ↑ Total Tokens ↓ LightMem 0.3308 0.4130 0.0718 0.2744 0.2875 825,896 MINDSET 0.3306 0.4293 0.0745 0.2806 0.2957 768,173 Nemori 0.3295 0.3085 0.0510 0.1794 0.1937 1,440,066 Vector RAG 0.3226 0.3904 0.0670 0.2675 0.2779 705,867 Long Context 0.2739 0.0153 0.0036 0.0082 0.0082 5,000,182 MemoryBank 0.2582 0.0678 0.0115 0.0453 0.0488 979,396 Table 6: Gemma-4-31B on the same complete LoCoMo-700 manifest. 5.3 Main results on MemoryAgentBench MAB appears difficult for every evaluated system, the leading ICL exact-match score stays only 0.024. MINDSET ties/leads all four reported measures. Its recommendation Recall@5 is 0.14 compared with MemoryBank’s 0.10 and a softer ICL token F1. The subset contains only 25 recommendation cases, and the five ICL groups remain challenging for all methods with MINDSET and Nemori getting the best results. We view MAB as evidence suggesting transfer across task type. 5.4 Ablating MINDSET Replacing energy-based assignment with nearest-schema merging and removing fragmentation pressure drop F1 by 0.0200 and 0.0237 respectively, showing the most impact. They support the central proposal: it matters not only whether relevant episodes can be found, but whether memory evolves into a useful number of coherent states. The other terms show smaller individual effects. Their retrieval columns except provenance expansion almost tie the full system as they affect how evidence is organized more than the presence of candidate episodes. 5.5 Cross-backbone transfer We additionally evaluate MINDSET using GLM-4.7 and Gemma-4-31B to show generalization across models. MINDSET remains best with GLM-4.7. With Gemma, LightMem leads by only 0.0002 F1 (0.3308 versus 0.3306), while MINDSET remains first on retrieval metrics. These results support model-agnostic operation - the memory layer runs without retraining and preserves its retrieval behavior although every generator does not convert that evidence into identical answers. 6 Conclusion Long-term conversational memory is not always about preserving the most text. We show that it can be handled by maintaining a changing state without erasing the path by which that state changed. MINDSET achieves this using immutable episodes, versioned schemas and minimum-energy transitions (reinforcement, supersession, contextual splitting and new memory). Across the frozen GPT-OSS experiments, this design produces the best observed LoCoMo answer and way better retrieval-ranking results. It also transfers to MAB results, although absolute metrics stay lower and difference is not significant. Ablations connect the gain to the structure of the proposal: energy-based assignment and controlled fragmentation matter more than simply increasing retrieval coverage. The cross-model backbone experiments across GLM and Gemma showed generalization without needing retraining. Limitations 1. Absolute MAB ICL scores are low and the recommendation split contains only 25 examples. These results therefore indicate relative performance under the evaluated protocol rather than broad MAB competence. 2. LightMem is evaluated through documented reproductions because all upstream trained artifacts were unavailable. MemoryBank additionally uses GPT-OSS in place of its original summarization model. 3. The transition coefficients and thresholds are hand-tuned engineering choices calibrated on held-out development data rather than scientifically proven optimal values. References Bae et al. (2022) S. Bae, D. Kwak, S. Kang, M. Y. Lee, S. Kim, Y. Jeong, H. Kim, S. Lee, W. Park, and N. Sung Keep me updated! memory management in long-term conversations. In Findings of the Association for Computational Linguistics: EMNLP 2022, Y. Goldberg, Z. Kozareva, and Y. Zhang (Eds.), Abu Dhabi, United Arab Emirates, p. 3769–3787. External Links: Link, Document Cited by: §2.3. Bai et al. (2024) Y. Bai, X. Lv, J. Zhang, H. Lyu, J. Tang, Z. Huang, Z. Du, X. Liu, A. Zeng, L. Hou, Y. Dong, J. Tang, and J. Li LongBench: a bilingual, multitask benchmark for long context understanding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), L. Ku, A. Martins, and V. Srikumar (Eds.), Bangkok, Thailand, p. 3119–3137. External Links: Link, Document Cited by: §2.1. Carbonell and Stewart (1999) J. Carbonell and J. Stewart The use of mmr, diversity-based reranking for reordering documents and producing summaries. SIGIR Forum (ACM Special Interest Group on Information Retrieval), p. . External Links: Document Cited by: §2.1. Chhikara et al. (2025) P. Chhikara, D. Khant, S. Aryan, T. Singh, and D. Yadav Mem0: building production-ready ai agents with scalable long-term memory. External Links: 2504.19413, Link Cited by: §2.2. Cormack et al. (2009) G. V. Cormack, C. L. A. Clarke, and S. Büttcher Reciprocal rank fusion outperforms condorcet and individual rank learning methods. Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval. External Links: Link Cited by: §2.1. Gutiérrez et al. (2025) B. J. Gutiérrez, Y. Shu, Y. Gu, M. Yasunaga, and Y. Su HippoRAG: neurobiologically inspired long-term memory for large language models. External Links: 2405.14831, Link Cited by: §1, §2.2. Hsieh et al. (2024) C. Hsieh, S. Sun, S. Kriman, S. Acharya, D. Rekesh, F. Jia, Y. Zhang, and B. Ginsburg RULER: what’s the real context size of your long-context language models?. External Links: 2404.06654, Link Cited by: §2.1. Hu et al. (2025) M. Hu, T. Chen, Q. Chen, Y. Mu, W. Shao, and P. Luo HiAgent: hierarchical working memory management for solving long-horizon agent tasks with large language model. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 32779–32798. External Links: Link, Document, ISBN 979-8-89176-251-0 Cited by: §2.3. Hu et al. (2026) Y. Hu, Y. Wang, and J. McAuley Evaluating memory in llm agents via incremental multi-turn interactions. External Links: 2507.05257, Link Cited by: §2.3, §4.1. Järvelin and Kekäläinen (2002) K. Järvelin and J. Kekäläinen Cumulated gain-based evaluation of ir techniques. ACM Trans. Inf. Syst. 20, p. 422–446. External Links: Link Cited by: §4.4. Jia et al. (2025) Z. Jia, Q. Liu, H. Li, Y. Chen, and J. Liu Evaluating the long-term memory of large language models. In Findings of the Association for Computational Linguistics: ACL 2025, W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 19759–19777. External Links: Link, Document, ISBN 979-8-89176-256-5 Cited by: §2.3. Kang et al. (2025) J. Kang, M. Ji, Z. Zhao, and T. Bai Memory OS of AI agent. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng (Eds.), Suzhou, China, p. 25961–25970. External Links: Link, Document, ISBN 979-8-89176-332-6 Cited by: §2.2, §2.3. Karpukhin et al. (2020) V. Karpukhin, B. Oguz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, and W. Yih Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), B. Webber, T. Cohn, Y. He, and Y. Liu (Eds.), Online, p. 6769–6781. External Links: Link, Document Cited by: §1, §2.1. LeCun et al. (2006) Y. LeCun, S. Chopra, R. Hadsell, A. Ranzato, and F. J. Huang A tutorial on energy-based learning. External Links: Link Cited by: §2.3. Lewis et al. (2021) P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rocktäschel, S. Riedel, and D. Kiela Retrieval-augmented generation for knowledge-intensive nlp tasks. External Links: 2005.11401, Link Cited by: §1, §2.1. Liu et al. (2024) N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang Lost in the middle: how language models use long contexts. Transactions of the Association for Computational Linguistics 12, p. 157–173. External Links: Link, Document Cited by: §1, §2.1. Ma et al. (2026) W. Ma, J. Nan, and W. Wu What deserves memory: adaptive memory distillation for LLM agents. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), M. Liakata, V. P. Moreira, J. Zhang, and D. Jurgens (Eds.), San Diego, California, United States, p. 34789–34812. External Links: Link, Document, ISBN 979-8-89176-390-6 Cited by: §1, §2.2, 5th item. Maharana et al. (2024) A. Maharana, D. Lee, S. Tulyakov, M. Bansal, F. Barbieri, and Y. Fang Evaluating very long-term conversational memory of LLM agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), L. Ku, A. Martins, and V. Srikumar (Eds.), Bangkok, Thailand, p. 13851–13870. External Links: Link, Document Cited by: §2.3, §4.1. Mcclelland et al. (1995) J. Mcclelland, B. Mcnaughton, and R. O’Reilly Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological Review 102, p. 419–457. External Links: Document Cited by: §2.3. Ong et al. (2025) K. T. Ong, N. Kim, M. Gwak, H. Chae, T. Kwon, Y. Jo, S. Hwang, D. Lee, and J. Yeo Towards lifelong dialogue agents via timeline-based memory management. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), L. Chiruzzo, A. Ritter, and L. Wang (Eds.), Albuquerque, New Mexico, p. 8631–8661. External Links: Link, Document, ISBN 979-8-89176-189-6 Cited by: §2.3. Packer et al. (2024) C. Packer, S. Wooders, K. Lin, V. Fang, S. G. Patil, I. Stoica, and J. E. Gonzalez MemGPT: towards llms as operating systems. External Links: 2310.08560, Link Cited by: §1, §2.2. Park et al. (2023) J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein Generative agents: interactive simulacra of human behavior. External Links: 2304.03442, Link Cited by: §1, §2.2. Reimers and Gurevych (2019) N. Reimers and I. Gurevych Sentence-BERT: sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), K. Inui, J. Jiang, V. Ng, and X. Wan (Eds.), Hong Kong, China, p. 3982–3992. External Links: Link, Document Cited by: §1, §2.1. Robertson and Zaragoza (2009) S. Robertson and H. Zaragoza The probabilistic relevance framework: bm25 and beyond. Found. Trends Inf. Retr. 3 (4), p. 333–389. External Links: ISSN 1554-0669, Link, Document Cited by: §2.1. Salama et al. (2025) R. Salama, J. Cai, M. Yuan, A. Currey, M. Sunkara, Y. Zhang, and Y. Benajiba MemInsight: autonomous memory augmentation for LLM agents. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng (Eds.), Suzhou, China, p. 33136–33152. External Links: Link, Document, ISBN 979-8-89176-332-6 Cited by: §2.3. Shinn et al. (2023) N. Shinn, F. Cassano, E. Berman, A. Gopinath, K. Narasimhan, and S. Yao Reflexion: language agents with verbal reinforcement learning. External Links: 2303.11366, Link Cited by: §1, §2.2. Tan et al. (2025a) H. Tan, Z. Zhang, C. Ma, X. Chen, Q. Dai, and Z. Dong MemBench: towards more comprehensive evaluation on the memory of LLM-based agents. In Findings of the Association for Computational Linguistics: ACL 2025, W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 19336–19352. External Links: Link, Document, ISBN 979-8-89176-256-5 Cited by: §2.3. Tan et al. (2025b) Z. Tan, J. Yan, I. Hsu, R. Han, Z. Wang, L. Le, Y. Song, Y. Chen, H. Palangi, G. Lee, A. R. Iyer, T. Chen, H. Liu, C. Lee, and T. Pfister In prospect and retrospect: reflective memory management for long-term personalized dialogue agents. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 8416–8439. External Links: Link, Document, ISBN 979-8-89176-251-0 Cited by: §2.3. Wang et al. (2024) Z. Wang, Z. Li, Z. Jiang, D. Tu, and W. Shi Crafting personalized agents through retrieval-augmented generation on editable memory graphs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Y. Al-Onaizan, M. Bansal, and Y. Chen (Eds.), Miami, Florida, USA, p. 4891–4906. External Links: Link, Document Cited by: §2.3. Wu et al. (2025) D. Wu, H. Wang, W. Yu, Y. Zhang, K. Chang, and D. Yu LongMemEval: benchmarking chat assistants on long-term interactive memory. External Links: 2410.10813, Link Cited by: §2.3. Xu et al. (2022) J. Xu, A. Szlam, and J. Weston Beyond goldfish memory: long-term open-domain conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), S. Muresan, P. Nakov, and A. Villavicencio (Eds.), Dublin, Ireland, p. 5180–5197. External Links: Link, Document Cited by: §2.3. Xu et al. (2025) W. Xu, Z. Liang, K. Mei, H. Gao, J. Tan, and Y. Zhang A-mem: agentic memory for llm agents. External Links: 2502.12110, Link Cited by: §2.2. Yao et al. (2023) S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao ReAct: synergizing reasoning and acting in language models. External Links: 2210.03629, Link Cited by: §2.2. Yu et al. (2026) Y. Yu, L. Yao, Y. Xie, Q. Tan, J. Feng, Y. Li, and L. Wu Agentic memory: learning unified long-term and short-term memory management for large language model agents. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), M. Liakata, V. P. Moreira, J. Zhang, and D. Jurgens (Eds.), San Diego, California, United States, p. 21457–21483. External Links: Link, Document, ISBN 979-8-89176-390-6 Cited by: §1, §2.2. Zacks et al. (2007) J. M. Zacks et al. Event Perception: A Mind-Brain Perspective. Note: Psychological Bulletin 133(2). Cited by: §2.3. Zhang et al. (2026) J. Zhang, C. Zhang, S. Chen, Z. Huang, P. Zheng, Z. Wang, P. Guo, F. Mo, S. Bae, J. Zou, J. Wei, and Y. Yang Lightweight LLM agent memory with small language models. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), M. Liakata, V. P. Moreira, J. Zhang, and D. Jurgens (Eds.), San Diego, California, United States, p. 12914–12929. External Links: Link, Document, ISBN 979-8-89176-390-6 Cited by: §1, §2.2, 4th item. Zhang et al. (2025) Y. Zhang, W. Yuan, and Z. Jiang Bridging intuitive associations and deliberate recall: empowering LLM personal assistant with graph-structured long-term memory. In Findings of the Association for Computational Linguistics: ACL 2025, W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 17533–17547. External Links: Link, Document, ISBN 979-8-89176-256-5 Cited by: §2.3. Zhong et al. (2023) W. Zhong, L. Guo, Q. Gao, H. Ye, and Y. Wang MemoryBank: enhancing large language models with long-term memory. External Links: 2305.10250, Link Cited by: §1, §2.2, 3rd item. Appendix A Appendix A.1 Reproducibility controls The random seed is 0. MINDSET uses nearest-schema k=6k=6, minimum schema size 2, direct and schema pools of 24 each, version-schema k=4k=4, continuity radius 2 with up to 16 anchors, new-schema threshold 0.43, reinforce threshold 0.58, stay margin 0.06, and switch margin 0.18. Common retrieval uses candidate k=40k=40, top-8 final evidence, and cross-encoder-only final selection. The shared bounded-evidence adapter uses (20,768,96,8,8000,3)(20,768,96,8,8000,3) for candidate memories, chunk size, overlap, selected chunks, token budget and per-memory chunk cap. A.2 Full statistical results We show complete statistical tests computed on our primary dataset, LoCoMo. All tests use paired per-question scores from frozen run artifacts. Answer F1 and evidence recall use all 700 questions; ranking metrics use the 696 questions with nonempty gold evidence, matching the aggregation used in the main results table. Both the bootstrap and the two-sided exact sign-flip permutation test operate on the ten conversation clusters, preserving all questions from a conversation as one dependent unit. We report 95% bootstrap confidence intervals and Holm-adjusted p-values within each metric family. Positive differences favor MINDSET. W-T-L denotes per-question wins, ties and losses. Mean differences are computed from unrounded per-question values. The cluster-level design guards against artificially narrow uncertainty estimates caused by treating questions from the same conversation as independent. Baseline Metric Mean Δ 95% CI W–T–L Permutation p Holm p Vector RAG Answer F1 +0.0473 [+0.0178,+0.0783] 202–367–131 0.02148 0.04297 Long Context Answer F1 +0.1489 [+0.1094,+0.1857] 295–313–92 0.00195 0.00977 Nemori Answer F1 +0.0547 [+0.0188,+0.1074] 217–325–158 0.00977 0.02930 LightMem Answer F1 +0.0256 [-0.0118,+0.0636] 175–380–145 0.24023 0.24023 MemoryBank Answer F1 +0.1305 [+0.0786,+0.1822] 284–318–98 0.00391 0.01562 Vector RAG Recall@8 +0.2377 [+0.1970,+0.2815] 220–453–23 0.00195 0.00977 Long Context Recall@8 +0.6127 [+0.5745,+0.6489] 485–202–9 0.00195 0.00977 Nemori Recall@8 +0.2082 [+0.1788,+0.2363] 243–384–69 0.00195 0.00977 LightMem Recall@8 +0.1758 [+0.1345,+0.2227] 173–488–35 0.00195 0.00977 MemoryBank Recall@8 +0.5646 [+0.5189,+0.6030] 460–221–15 0.00195 0.00977 Vector RAG Precision@8 +0.0386 [+0.0324,+0.0451] 220–453–23 0.00195 0.00977 Long Context Precision@8 +0.1020 [+0.0946,+0.1088] 485–202–9 0.00195 0.00977 Nemori Precision@8 +0.0374 [+0.0301,+0.0437] 243–384–69 0.00195 0.00977 LightMem Precision@8 +0.0273 [+0.0205,+0.0350] 173–488–35 0.00195 0.00977 MemoryBank Precision@8 +0.0950 [+0.0899,+0.0996] 460–221–15 0.00195 0.00977 Vector RAG MRR +0.2897 [+0.2594,+0.3218] 348–299–49 0.00195 0.00977 Long Context MRR +0.5491 [+0.5211,+0.5796] 492–193–11 0.00195 0.00977 Nemori MRR +0.3045 [+0.2725,+0.3327] 368–238–90 0.00195 0.00977 LightMem MRR +0.1207 [+0.0915,+0.1500] 155–492–49 0.00195 0.00977 MemoryBank MRR +0.5134 [+0.4823,+0.5439] 475–211–10 0.00195 0.00977 Vector RAG nDCG@8 +0.2674 [+0.2356,+0.3023] 364–281–51 0.00195 0.00977 Long Context nDCG@8 +0.5371 [+0.5067,+0.5674] 492–193–11 0.00195 0.00977 Nemori nDCG@8 +0.2724 [+0.2517,+0.2944] 383–216–97 0.00195 0.00977 LightMem nDCG@8 +0.1361 [+0.1029,+0.1710] 183–454–59 0.00195 0.00977 MemoryBank nDCG@8 +0.4990 [+0.4647,+0.5312] 474–205–17 0.00195 0.00977 Vector RAG Evidence recall +0.2363 [+0.1959,+0.2797] 220–457–23 0.00195 0.00977 Long Context Evidence recall +0.4111 [+0.3673,+0.4575] 374–264–62 0.00195 0.00977 Nemori Evidence recall -0.0019 [-0.0303,+0.0272] 131–426–143 0.91602 1.00000 LightMem Evidence recall +0.1748 [+0.1343,+0.2202] 173–492–35 0.00195 0.00977 MemoryBank Evidence recall +0.0198 [-0.0882,+0.1254] 190–317–193 0.73242 1.00000 Table 7: Complete main LoCoMo significance tests: MINDSET minus each baseline. Ranking metrics use the 696 questions with nonempty gold evidence; answer F1 and evidence recall use all 700 questions. Mean differences are computed before rounding. Bold adjusted values satisfy p<0.01p<0.01. Interpretation. The strongest inferential result is retrieval ranking on LoCoMo. MINDSET significantly improves Recall@8, MRR and nDCG@8 over every baseline after Holm correction, including LightMem and Nemori. Its answer-F1 advantage is significant at p<0.05p<0.05 against Vector RAG, Long Context, Nemori and MemoryBank. However, the positive gain over LightMem is not statistically confirmed. A.3 Parameter sensitivity The frozen LoCoMo-700 MINDSET run ingests 5,882 dialogue episodes from the ten histories needed by those 700 questions. These episodes are memory input, not additional evaluation questions. The controller produces 179 schemas: 168 active and 11 historical. A schema contains 32.26 support episodes on average, or 32.86 provenance episodes when exceptions are included. Structural quantity Count Share / mean Ingested dialogue episodes 5,882 – Schemas 179 – Active schemas 168 93.85% Historical schemas 11 6.15% Support memberships 5,775 32.26/schema Exception memberships 107 0.60/schema All provenance memberships 5,882 32.86/schema Table 8: Frozen MINDSET structure for the exact LoCoMo-700 paper run. We vary one parameter at a time around its frozen value while keeping the exact 700-question manifest and all retrieval settings fixed. As in the main table, retrieval metrics average the 696 questions with annotated gold evidence; the four unannotated questions remain in the manifest but are excluded from ranked retrieval aggregation. This diagnostic is local and makes no answer-model or judge calls. Because these parameters act during memory construction, we report retrieval and structural behavior rather than reuse answers generated from a different evidence set. The neighborhood is flat. Across all rows, Recall@8 spans 0.0015, Precision@8 0.0002, MRR 0.0014, and nDCG@8 0.0012. Contradiction-weight and switch-margin changes do not alter a transition in this range; fragmentation moves at most one episode between new and split and changes the schema count by at most one. This supports local robustness, not uniqueness A.4 Answer prompt The system template is: ⬇ Answer every question using only the supplied memory evidence. Return one strict JSON object with an ‘answers‘ array. Each array item must contain exactly ‘question_id‘ and a concise ‘answer‘. Preserve every question ID exactly. Each question contains its own isolated ‘evidence‘ array; evidence from another question is out of scope. Exact names, numbers, dates, colors, attributes, and list members must be explicitly supported by that question’s grounding evidence. If the evidence does not support an answer, answer exactly ‘No information available.‘; do not guess. The user payload is canonical compact JSON with an array of objects containing question_id, question, and that question’s isolated ranked evidence. Keys are sorted. Responses must parse as one JSON object and preserve every ID; missing or malformed items trigger bounded retries. A.5 MINDSET transition pseudocode ⬇ for episode e in chronological conversation: if no active schema exists: create_new(e); continue target = max_compatibility(nearest_active_schemas(e, k=6)) signals = K, p, temporal tau, contextual kappa, recurrence r, explicit_change x energies = exact terms defined in Sections 3.2--3.3 if K < 0.43 and topic_match == 0 and dense < 0.35: action = new else: feasible = reinforce, new if p >= 0.35: feasible += supersede if kappa >= 0.12: feasible += split if K >= 0.43: feasible -= new if p < 0.15 and kappa < 0.12 and K >= 0.58: action = reinforce elif p >= 0.60 and r == 0 and x == 0: action = reinforce_as_exception elif feasible == reinforce: action = reinforce else: b = lowest_energy(feasible - reinforce) margin = 0.18 * (1 - 0.6*r) if tau > kappa and tau > 0.45: margin *= 0.55 delta = E(reinforce) - E(b) if delta > margin: action = b elif delta <= 0.06: action = reinforce else: action = b apply(action); retain raw episode provenance A.6 Retrieval and evidence construction ⬇ question -> classify current / historical / contextual / generic -> direct dense + BM25 episode candidates -> relevant active/historical schemas -> provenance and continuity expansion -> local MiniLM cross-encoder reranking -> up to 20 memory candidates -> 768-token passages, 96-token overlap -> local dense passage scoring -> greedy top-8 under 8,000 tokens and <=3 passages/source -> provenance-preserving evidence JSON -> frozen answer model A.7 Ablation definitions • Nearest-schema merge: replace minimum-energy action selection with assignment to the nearest schema. • No contextual term: remove contextual transition potential. • No temporal term: remove temporal transition potential. • No hysteresis: remove stay/switch stabilization and isolated-contradiction holding behavior. • No fragmentation penalty: set the fragmentation energy contribution to zero. • No provenance expansion: retrieve through the selected memory units without expanding schema provenance and continuity anchors. A.8 Metric implementation notes LoCoMo answer F1 is deterministic and category aware. MAB ICL exact match parses the first answer line and removes an Answer: prefix. Recommendation scoring canonicalizes movie entities with the benchmark entity map and edit distance before Recall@5. Supplementary MAB ICL F1 selects the best normalized-token match among valid references; supplementary recommender F1 uses the canonicalized top-five set. No LLM judge is used for any number in the tables. A.9 Disclosure of unavailable and substituted components LightMem’s unpublished LoRA selector and offline consolidation model are substituted to support reproduction. MemoryBank uses the shared GPT-OSS model for its summary calls. These choices preserve experimental control but mean the corresponding rows should be read as results for documented reproductions, not exact reruns of every upstream software dependency. A.10 Development set We reserve a separate 300-question LoCoMo development set for hand-tuning MINDSET’s compatibility coefficients, energy weights, transition thresholds, hysteresis margins and retrieval-coverage parameters before the primary evaluation. The set is disjoint from the frozen LoCoMo-700 manifest at the question-ID level and contains 75 examples from each available category (categories 1, 2, 4, and 5); category 3 has no remaining examples because it is fully represented in the primary manifest. Selection is done after excluding every question in the primary 700 LoCoMo samples used for the main paper. Questions are sorted by the SHA-256 key, the first 75 questions in each available category are selected and the selected records are sorted globally by the same key. No primary-test question is used for parameter selection. The complete manifest, including exact question IDs, source indices, category labels and source-manifest hashes will be released alongside the codebase upon publication to make development selection and parameter tuning fully reproducible.