PaperlayerView exact revision on arXiv

Source-linked AI summary

PreCoMem: Predictive Cognitive Memory for Self-Evolving Long-Term Dialogue Agents

Jian Zhong, Zeyu Liu, Pingchuan Cao, Rongduo Han, Shunye Tang, Guohuan Xie, Mingyuan Qin, Ran Ji, Xiao Liang, Wei Wang, Haining Zhang

arXiv:2610.29467v1cs.AI

TL;DR

Long-horizon dialogue agents must preserve useful user memories while adapting to genuine changes amid sparse, noisy, and uncertain evidence. PreCoMem uses confidence-weighted beliefs and a reliability-adjusted surprisal gate to choose among reinforcement, provisional hypotheses, and correction. It reports strong benchmark accuracy while keeping memory bounded over a thousand turns.

  • Problem

    Long-horizon agents need to preserve stable user traits while adapting to changing preferences and life facts, despite sparse cues, noisy retrieval, and rarely explicit correction.

    Source: §1 · p. 2

  • Method

    PreCoMem retrieves confidence-weighted beliefs and uses reliability-adjusted surprisal with a two-threshold gate to select MAINTAIN, PROFILE, or CORRECT updates.

    Source: §1 · p. 2, §2.4 · p. 4

  • Results

    PreCoMem reports the best overall accuracy across LoCoMo, LongMemEval-S, PersonaMem-v2, and DiMoBench while keeping memory bounded over a thousand turns.

    Source: §4 · p. 10

  • Takeaways & Limitations

    Buffering ambiguous cues as low-confidence hypotheses is presented as the central design choice for staying consistent under noise without freezing under genuine drift.

    Source: §4 · p. 10

Abstract

from arXiv · show

Long-horizon dialogue agents need to keep user memories stable enough for personalization while adapting when the user genuinely changes. Existing memory systems often append every observation or overwrite old entries by heuristic rules, which makes transient noise hard to distinguish from real belief drift. We introduce PreCoMem, a memory consolidation framework inspired by predictive coding that represents memory as confidence-weighted beliefs about the user's state. Its core signal, Effective Surprisal, combines semantic deviation and contradiction evidence, then downweights this signal when retrieval is diffuse and unreliable. According to effective surprisal, a two-threshold gate maps each turn to one of three deterministic updates: MAINTAIN reinforces supported beliefs, PROFILE stores ambiguous cues as low-confidence hypotheses, and CORRECT softly decays reliably contradicted facts. The PROFILE stage acts as an explicit waiting room, preventing premature commitment to noisy evidence. Experiments on LoCoMo, LongMemEval-S, and PersonaMem-v2 show consistent state-of-the-art accuracy; our new DiMoBench confirms that PreCoMem adapts to non-stationary belief drift without over-reacting to noise. The code is available at https://github.com/ZhongWishing/PreCoMem.

1 Introduction

Long-horizon dialogue memory must preserve stable user traits while adapting to genuine change despite sparse cues, noisy retrieval, and rarely explicit correction. PreCoMem uses reliability-adjusted surprisal to route updates among reinforcement, provisional hypotheses, and correction.

1 Introduction

  • Long-horizon memory must balance preserving stable user traits with adapting to belief drift, despite sparse signals, noisy retrieval, and rarely explicit correction.The paper calls this the memory stability–plasticity problem and identifies premature correction and old-memory pull as opposing failures.Source: §1 · p. 2
  • Existing systems often use recency overwriting, fixed-threshold importance, or one-shot appending, which cannot reliably distinguish persistent drift from transient noise.Source: §1 · p. 2
  • Predictive coding motivates reinforcing events consistent with memory and treating surprising events as possible learning or correction, while accounting for retrieval reliability.The paper presents this as a design principle, not a claim that dialogue memory is a fully specified probabilistic model.Source: §1 · p. 2
  • PreCoMem retrieves confidence-weighted beliefs, computes reliability-adjusted Effective Surprisal, and routes turns to MAINTAIN, PROFILE, or CORRECT.PROFILE buffers ambiguous evidence as low-confidence hypotheses; CORRECT softly decays contradicted facts.Source: §1 · p. 2
  • PreCoMem outperforms the strongest contemporary baselines across its evaluated benchmarks while keeping effective memory stable across 1050 turns.Source: §1 · p. 2

2 PreCoMem

PreCoMem combines confidence-weighted belief retrieval with a deterministic memory-evolution loop. Its reliability-adjusted surprisal gate distinguishes consistent, ambiguous, and reliably contradictory evidence, while beliefs remain editable and hypotheses provisional until supported.

2.1 Framework overview

  • The framework reuses recalled beliefs in a read–write loop, with continuous confidence weights and PROFILE hypotheses distinguishing it from discrete commit or expiry decisions.PROFILE lets weak evidence accumulate across turns before commitment or discard.Source: §2.1 · pp. 2–3

2.2 Memory primitives

  • The belief store holds editable facts and hypotheses with confidence weights, while an episodic cache preserves immutable dialogue traces and source pointers.Hypotheses can gain support, decay when unsupported, and become facts after crossing a promotion threshold.Source: §2.2 · pp. 3–4

2.3 Memory leveraging: evidence recall and response generation

  • Memory Leveraging ranks beliefs using semantic similarity, lexical overlap, and topic relevance, modulated by confidence, then uses Top-K evidence to condition response generation.Linked episodic traces provide provenance context but do not participate in Top-K ranking.Source: §2.3 · p. 4

2.4 Memory evolution: regime-gated belief updating

  • Retrieval reliability is estimated from score concentration: diffuse recall downweights deviation signals, whereas concentrated recall preserves them for correction.The reliability coefficient can defer an update but cannot trigger one by itself.Source: §2.4 · p. 4
  • Two thresholds route low surprisal to MAINTAIN, medium surprisal to PROFILE, and high surprisal to CORRECT.The main experiments fix the thresholds at 0.2 and 0.6.Source: §2.4 · p. 5
  • MAINTAIN strengthens supported beliefs, while CORRECT records a new fact and selectively decays contradicted retrieved beliefs.Correction severity is item-specific; non-contradicted items are essentially unchanged.Source: §2.4 · p. 5
  • PROFILE writes novel but not decisively contradictory observations as low-confidence hypotheses that need repeated support to become facts.Unsupported hypotheses decay over time, while supporting updates reset their decay clock.Source: §2.4 · p. 5

3 Experiments

Across long-term memory benchmarks, PreCoMem leads the reported overall comparisons and recovers from simulated belief drift. Ablations, retrieval-budget tests, and memory analyses examine the update operators, retrieval reliability, and storage behavior underlying these results.

3.1 Experimental setup

  • The evaluation covers LoCoMo, LongMemEval-S, PersonaMem-v2 at 32k and 128k contexts, and controlled DiMoBench drift; answers are scored by three LLM judges.DiMoBench comprises five 350-turn dialogues and DiMoBench-XL extends to 1050 turns with three sequential drifts.Source: §3.1 · pp. 6–7, §3.3 · p. 9

3.2 Main results

  • PreCoMem achieves the best overall accuracy on LoCoMo and LongMemEval-S and leads on PersonaMem-v2; with the answer model fixed, its smaller-backbone memory construction wins all twelve Figure 3 comparisons.On LoCoMo, its margins over TiMem are +1.69 with gpt-4o-mini and +1.02 with gpt-4.1; on LongMemEval-S it leads EverMemOS by +3.23 under gpt-4.1. On PersonaMem-v2 Open-Ended it leads EverMemOS by +3.37 at 32k and +2.89 at 128k.Source: §3.2 · pp. 7–8, Figure 3 · p. 7

3.3 Ablation studies and belief-drift adaptation

  • Removing the effective-surprisal gate causes the largest reported LoCoMo ablation drop, from 73.57 to 31.46; removing confidence weighting or retrieval-reliability adjustment also substantially reduces accuracy.End-to-end LLM write decisions lower accuracy by 2.75 / 3.15 points on LoCoMo / LongMemEval-S and raise cross-seed standard deviation 20× on LoCoMo.Source: §3.3 · pp. 8–9
  • The component ablations hold backbone, prompts, and hyperparameters fixed, while the DiMoBench suite tests gradual attitude and preference drift amid keyword-matched chatter and entity confounders.The base suite has five independent 350-turn dialogues with 1,000 QA pairs total; DiMoBench-XL has five 1050-turn dialogues with three drifts and 2,100 QA pairs.Source: §3.3 · pp. 8–9
  • On DiMoBench, PreCoMem recovers fastest after the change point and stabilizes at 84%, while EverMemOS, TiMem, and O-Mem plateau at 78–80% and Mem0 ends at 54%.Removing PROFILE caps the asymptote at 72%; removing CORRECT delays recovery from turn 250 to 320 and lowers the plateau to 77%; removing MAINTAIN ends at 54%.Source: Figure 4 · p. 8, §3.3 · p. 9

3.4 Hyperparameter analysis

  • On LoCoMo, accuracy peaks at k=10; larger retrieval budgets yield diminishing returns while lowering retrieval reliability and dampening effective surprisal.A retrieved item can also spuriously appear contradictory under the max-based contradiction component, motivating the fixed k=10 setting.Source: §3.4 · p. 9, Figure 5 · p. 9

3.5 Memory scalability

  • Under long-horizon drift, effective memory saturates by turn 200 at 25.8–28.5 on the representative persona, while raw memory grows roughly linearly.Deprecated facts occupy 1.0% of Top-10 slots on average after 1050 turns and three drifts. On LoCoMo, optional pruning stores fewer items than both baselines while retaining an accuracy lead of +2.6 and +3.4 overall points.Source: §3.5 · pp. 9–10

4 Conclusion

The conclusion frames PreCoMem as confidence-weighted beliefs revised through reliability-adjusted surprisal and three update operators. The supplementary discussion situates this write-path design against prior systems and describes an on-policy approximation, retrieval signals, and memory-item construction.

4 Conclusion

  • PreCoMem represents conversational memory as confidence-weighted beliefs and uses MAINTAIN, PROFILE, and CORRECT to reinforce, buffer, or softly decay evidence; its evaluation is limited to English dialogue benchmarks.The conclusion presents continuous confidence and explicit hypothesis lifecycles as a useful baseline for long-term memory systems.Source: §4 · p. 10

A Related work

  • Prior memory systems improve storage organization or retrieval, but the related-work discussion identifies write decisions under gradual drift and adversarial retrieval interference as an unisolated benchmark gap.DiMoBench and DiMoBench-XL are introduced as controlled drift benchmarks addressing that gap.Source: A Related work · pp. 15–16

B.1 From expected predictive loss to a per-turn on-policy loss

  • The appendix motivates a per-turn on-policy loss as a Monte Carlo estimate of expected effective surprisal under a fixed response policy, then approximates long-horizon updates with a regularized surrogate.The surrogate includes an edit/stability cost and a capacity penalty; hypotheses below confidence 0.1 are pruned, whereas outdated facts are softened by confidence decay rather than direct deletion.Source: §B.1 · p. 16
  • Low-confidence facts remain stored because confidence-weighted retrieval makes them nearly inactive while retaining provenance and links that may help recover a trait if it resurfaces.For extreme horizons, the text suggests timestamped time-to-live as an engineering safeguard for raw storage.Source: §B.1 · p. 17

B.2 Retrieval scoring and surprisal component definitions

  • Retrieval scoring combines normalized lexical relevance, topical relevance, and dense evidence; topic labels are assigned from the belief text using topic prototypes.Dense embeddings use belief text alone, while entity and topic metadata contribute through their separate terms.Source: §B.2 · pp. 17–18
  • The surprisal components include semantic deviation from a score-weighted evidence embedding and a contradiction probability supplied by a calibrated NLI verifier.The verifier’s contradiction probability is obtained by temperature-scaling its entailment, neutral, and contradiction logits.Source: §B.2 · p. 17
  • When the belief store is empty, retrieval supplies no evidence, so the first observations are conventionally stored as provisional PROFILE hypotheses for later reinforcement, correction, or decay.Source: §B.2 · p. 18

B.3 Operator and gate derivations

PreCoMem implements consolidation and revision through confidence-weighted updates, a two-threshold gate, and a hypothesis lifecycle. Its per-turn pipeline combines deterministic calculations and updates with regime-specific LLM sub-agents.

B.3 Operator and gate derivations

  • MAINTAIN reinforces supported beliefs with a gradient-descent step that shrinks the residual confidence gap by a factor of (1 − η).For weights in [0, 1] and η in (0, 1], the updated weight remains in [0, 1].Source: §B.3 · p. 18
  • CORRECT applies a KL-proximal update whose scalar-weight form matches the paper’s soft-decay rule, with contradiction severity supplied per retrieved item.The two-threshold gate is described as selecting actions by minimizing on-policy risk plus action cost.Source: §B.3 · p. 18
  • The two-threshold gate treats low, medium, and high effective surprisal as consistent, ambiguous, and reliably contradictory evidence, respectively.The thresholds (0.2, 0.6) are calibrated empirically; the design abstractions motivate ordered boundaries rather than being estimated end-to-end.Source: §B.3 · pp. 18–19
  • PROFILE initializes hypotheses by averaging a class-dependent prior with the agent’s confidence clipped to that class’s band, which prevents any single agent output from committing a fact.The class band sets the main semantics, while the within-band value fine-tunes the starting confidence.Source: §B.3 · p. 20
  • Unprotected hypotheses without support decay from their last evidence-bearing update and are pruned below weight 0.1; at γ = 0.05, never-reinforced hypotheses are pruned after eight unsupported turns.Reinforcement resets the decay window, while explicit-event hypotheses are protected.Source: §B.3 · p. 20

B.4 Per-turn pipeline (hybrid execution)

  • The hybrid pipeline computes retrieval reliability and effective surprisal deterministically, uses an LLM sub-agent for regime-specific targets, then applies closed-form updates and lifecycle rules symbolically.The sub-agents select items and strengths; gating, reinforcement, soft decay, and promotion are deterministic.Source: §B.3 · p. 19, §B.4 · p. 20

C.1 Third-party benchmarks: AMemGym and LoCoMo-Plus

On two third-party benchmarks, PreCoMem outperforms reported comparison systems and shows operator-specific effects on state-change probes. Additional analyses report gains when delegating memory construction to a smaller backbone, alongside bounded response-model token use.

C.1 Third-party benchmarks: AMemGym and LoCoMo-Plus

  • PreCoMem leads AMemGym’s strongest reported system, AWE, by +0.028 Overall and +0.033 Memory.Table 3 covers the official 20-profile, 11-period setting and compares systems using gpt-4.1-mini; PreCoMem results average three seeds.Source: §C.1 · pp. 20–21, Table 3 · p. 21
  • AMemGym operator ablations produce distinct probe signatures: removing PROFILE mainly lowers first-write success and changed-state accuracy, while removing CORRECT raises stale-state error and slows recovery.Removing MAINTAIN degrades stable-state accuracy; paired 95% confidence intervals for Full minus each variant exclude zero on every metric.Source: §C.1 · p. 21, Table 4 · p. 21
  • Static QA benchmarks provide a single end score and cannot attribute an operator’s effect across stable, change-point, and recovery phases.The paper therefore reports operator ablations on AMemGym and DiMoBench, with route-matched controls elsewhere.Source: §C.1 · p. 21
  • On LoCoMo-Plus, PreCoMem reaches 22.4 ± 0.8 overall constraint-consistency accuracy, above full-context GPT-4o (21.05) and the listed memory baselines.Gains over the w/o PROFILE control are largest on goal (+5.0) and state (+4.8), while causal (+1.1 ± 0.5) is within seed noise.Source: §C.1 · p. 22, Table 5 · p. 22

C.2 Backbone-generalization protocol and token usage

  • Across all twelve model-and-dataset comparisons, a one-tier-smaller PreCoMem memory system beats the answer model’s native memory while holding the answer model fixed.The reported comparison changes memory construction and retrieval, not the answer model.Source: §C.2 · p. 22
  • PreCoMem’s response-model input budget ranges from 3.1k tokens on LoCoMo to 3.8k on LongMemEval-S, and grows by under 9% when PersonaMem-v2 context quadruples.No baseline is both cheaper and more accurate on any benchmark; MemOS and EverMemOS use 0.6–1.4k fewer tokens per query at lower accuracy.Source: §C.2 · p. 22

C.3 Extended results and judge validation

Extended evaluations report a LongMemEval-S advantage under gpt-4o-mini and analyze which question types account for drift-related gains. The paper also reports judge-human agreement and lifecycle statistics on LoCoMo.

C.3 Extended results and judge validation

  • 84.28 ± 0.05 overall accuracy on LongMemEval-S under gpt-4o-mini leads EverMemOS by +1.96, with PreCoMem also leading on SSA, SSP, MS, and KU.EverMemOS leads on SSU and TiMem leads on temporal reasoning (TR).Source: §C.3 · p. 22, Table 7 · p. 23
  • The primary judge agrees with expert human re-annotation at 96.9% overall across LoCoMo, LongMemEval-S, and DiMoBench, with Cohen’s κ from 0.81 to 0.96 across question types.A cross-method spot-check on EverMemOS and TiMem outputs reports agreement within a ±0.7-point band of the pooled PreCoMem agreement.Source: §C.3 · p. 23, Table 8 · p. 24
  • On LoCoMo, removing retrieval-reliability adjustment increases hypothesis creation but reduces promotion precision, while removing PROFILE weakens gradual adaptation.Without the hypothesis lifecycle, medium-surprisal information is written as facts directly, making the lifecycle metrics undefined.Source: §C.3 · p. 23, Table 9 · p. 24
  • The paper attributes drift-related gains to KU, SSP, cross-session MS, and LoCoMo Open-Domain, while interpreting SSA gains as selective memory use through confidence-weighted retrieval.All 78 LongMemEval-S KU questions pair an old answer value with its superseding value; Open-Domain questions often require inferring latent states from multiple evidence pointers or sessions.Source: §C.3 · p. 24

C.4 Component, operator, and control ablations

The ablations test the surprisal signals, lifecycle, operators, and routing controls. They report distinct costs for removing components and show that direct LLM writes are less accurate and less stable than the constrained update path.

C.4 Component, operator, and control ablations

  • Figure 6 traces hypotheses from low confidence toward promotion, or decay and pruning when unsupported; reliably contradicted committed facts instead decay while remaining for provenance.The horizontal axis encodes confidence from low through the promotion threshold to stable high confidence.Source: Figure 6 · p. 25
  • Removing the contradiction signal costs −10.08 points on LoCoMo and −11.57 on LongMemEval-S, roughly double the drops from removing the semantic-gap signal.Both signals contribute; the paper attributes dconf’s larger contribution to detecting reliable negation.Source: §C.4 · p. 25, Table 10 · p. 25
  • On DiMoBench, removing PROFILE lowers the final accuracy ceiling to 72% versus 84% full, removing CORRECT slows recovery, and removing MAINTAIN ends at 54% without stabilization.The benchmark separates stable pre-change, mixed change-point, and stable post-change phases.Source: §C.4 · p. 26
  • Without PROFILE, a single-threshold gate reaches 69.92 / 81.16, down 3.65 / 3.12 against Full; random routing loses 11.74 / 11.34 with operator frequencies held fixed.The factorial controls report that removing lifecycle confidence alone costs 20.69 / 22.63 points, versus 1.77 / 0.72 for removing read-time weighting alone.Source: §C.4 · p. 27, Table 12 · p. 27
  • Direct LLM writes are lower in accuracy in every seed (−2.75 on LoCoMo, −3.15 on LongMemEval-S) and have 20× higher cross-seed standard deviation on LoCoMo.The comparison also reports slightly higher per-query token use; unconstrained writes sometimes emit missing target IDs, self-contradictory operations, or invalid JSON.Source: §C.4 · p. 27, Table 13 · p. 27
  • Per-item routing improves agreement on 145 genuinely mixed turns by +5.5 points and their question accuracy by +3.5, but yields only +0.3 points overall at higher write cost.The mixed turns are 1.61% of all annotated turns; the per-item variant uses +31% write calls and +14% write tokens.Source: §C.4 · p. 28

C.5 Hyperparameter sensitivity and adaptive variants

PreCoMem’s retrieval budget and update thresholds affect both evidence coverage and the reliability of memory updates. Adaptive thresholds improve benchmark accuracy, while the fixed default remains locally robust but is not claimed to be optimal for every sample.

Sensitivity to retrieval budget k

  • On LoCoMo, overall accuracy rises from 75.2 at k=1 to 85.06 at k=10, then falls to 82.5 at k=20.The section identifies k=10 as a robust default under a moderate compute budget.Source: §C.5 · pp. 28–29

Mechanism-level analysis

  • Larger k can flatten retrieval scores and lower ρt, making CORRECT harder to trigger, while a max-contradiction statistic becomes more exposed to spurious mismatches.These effects explain how accuracy can decline at large retrieval budgets despite broader coverage.Source: §C.5 · p. 29

Sensitivity to gating thresholds

  • The deployed thresholds (0.20, 0.60) sit near the center of a high-performing plateau, with immediate neighbors within 1.3–2.6 accuracy points of the peak.The study sweeps 25 threshold pairs on LoCoMo Sample 1; accuracy degrades smoothly toward the grid corners.Source: Figure 8 · p. 30, §C.5 · p. 30
  • The fixed pair is a stable, interpretable operating point for controlled comparisons, not a universally optimal choice for every LoCoMo sample.Dynamic threshold adaptation adds computation and can confound controlled evaluation by entangling behavior with an evolving schedule.Source: §C.5 · p. 30

Adaptive thresholding for personalization

  • Adaptive thresholds improve overall accuracy by +0.98 on gpt-4o-mini and +1.15 on gpt-4.1 on LoCoMo, and lift LongMemEval-S from 84.28 to 85.36 on gpt-4o-mini.The largest reported gains are +3.92 on LoCoMo Open-Domain, +1.62 on LongMemEval-S KU, and +1.95 on SSP.Source: Table 15 · p. 31, §C.5 · p. 31
  • The adaptive variant maps a user stability index to thresholds: stable users receive higher thresholds, while dynamic users receive lower thresholds, with exponential smoothing to limit oscillation.The proposed mapping reads existing summary statistics and adds no trainable parameters.Source: §C.5 · pp. 30–31

C.6 DiMoBench: construction and extended-horizon stress test

DiMoBench targets write-path adaptation under implicit, gradual belief drift, and DiMoBench-XL extends the stress test to longer dialogues with repeated drifts. Across 1050 turns, accuracy declines modestly while retrieval confidence remains high, and memory dynamics stay stable across personas.

DiMoBench construction and characteristics

  • DiMoBench uses five 350-turn dialogues with stable behavior, gradual implicit changes, reinforced post-change behavior, and a low-frequency maintenance phase, plus confounders and four QA reasoning types.Its questions avoid explicit sentiment keywords and require reasoning from behavioral cues; results are averaged across five personas.Source: §C.6 · pp. 32–33

DiMoBench construction and motivation

  • DiMoBench fills a write-path evaluation gap by testing whether systems maintain, update, or correct beliefs under continuous non-stationarity, implicit behavioral cues, and adversarial retrieval interference.The benchmark is designed to expose old-memory pull and premature correction, which read-path benchmarks do not isolate together.Source: §C.6 · p. 32

Six-stage construction pipeline

  • A six-stage pipeline separates GPT-4o generation from Gemini-2.5-Pro verification, followed by stratified human review for quality control.Verification checks six QA criteria, while 20% of the data receives stratified human review after annotators reach κ ≥0.75 agreement.Source: §C.6 · pp. 32–33, Table 16 · p. 32, Figure 9 · p. 33

DiMoBench-XL construction

  • DiMoBench-XL comprises five 1050-turn dialogues with three implicit drifts per persona and 2,100 evaluation QA pairs, achieving a 96.4% average verifier ACCEPT rate.The suite is verified with the same six-criterion protocol as DiMoBench.Source: §C.6 · p. 34

Extended-horizon stress test

  • On DiMoBench-XL, effective memory reaches a 25.8–28.5 plateau by turn 200, dips by at most 1.6 points across personas at drift, and recovers within about 50 turns.Nactive remains within 16–19 after turn 200, indicating that each drift replaces rather than accumulates the active belief set.Source: §C.6 · p. 34
  • Across 1050 turns and three drifts, DiMoBench-XL overall accuracy decays only 2.3 points, from 84.3% to 82.0% on average, while Top-10 retrieval confidence stays at or above 0.78.Accuracy standard deviations across personas remain below 1.0 point in every time window.Source: §C.6 · p. 34

C.7 Memory dynamics: footprint, contamination, and retention

The memory analysis finds that effective memory remains bounded over long dialogues, deprecated facts contribute little to retrieval, and more aggressive fact pruning trades smaller storage for reduced accuracy.

Retrieval contamination

  • Even at k=25, deprecated FACTs occupy only 2.6% of retrieval slots and contribute 0.94% of score mass; at k=10, the figures are 1.0% and 0.58%.Retrieved-item average confidence remains between 0.71 and 0.84 across settings.Source: §C.7 · p. 35, Table 20 · p. 36

Memory scalability and retention

  • Across 1050 turns and three drifts, effective memory converges to a stable plateau and deprecated FACT contamination in Top-10 retrieval stays at or below 1.3% in every window.The reported trajectory averages five personas; per-persona deviations are at most 0.5 on end effective memory.Source: Table 19 · p. 36

FACT pruning ablation

  • Pruning FACTs below 0.05 cuts raw memory from 287 to 168 items on LoCoMo with a 0.06-point Overall loss, while pruning below 0.10 loses 0.5 points on LoCoMo and 1.1 on DiMoBench.The authors retain the default provenance policy and recommend the more aggressive A1 variant for storage-bounded deployments.Source: §C.7 · pp. 35–36

Cross-system memory footprint

  • PreCoMem-A1 stores 168 items on the ten LoCoMo dialogues, fewer than O-Mem’s 232 and EverMemOS’s 264, while delivering +2.6 and +3.4 Overall points respectively.O-Mem and EverMemOS use discrete interaction decisions, so they have no per-item confidence and their effective count is undefined.Source: §C.7 · p. 36

C.8 Retrieval dependence and reliability calibration

Update correction depends on retrieving the belief that must change, and retrieval reliability calibrates how strongly surprisal influences the gate. The analyses quantify retrieval misses, correction delays, and cases where new-topic cues enter the hypothesis lifecycle.

Retrieval dependence

  • A retrieved target is necessary for same-turn correction: when the belief requiring revision is absent from the retrieved set, the write path cannot detect its contradiction or trigger CORRECT.The target belief is retrieved in 89–95% of update-due events.Source: §C.8 · p. 36

Belief-update recall and failure decomposition

  • Recall@10 covers 89–95% of update-due events, leaving a 5–11% blind spot; pooled across four benchmarks, 10.7% of events fail, with retrieval misses the largest single source.Question accuracy on retrieved-target events exceeds missed-event accuracy by 16–22 points, but forcing retrieval closes only 26–32% of the hit–miss gap.Source: §C.8 · pp. 36–37, Table 23 · p. 38
  • Missed update events do not always prevent later adaptation: 20% of missed events are corrected by subsequent cues or hypothesis promotion.The reported count is 147 of 718 missed events.Source: §C.8 · p. 37

Reliability calibration

  • On AMemGym, precision, contradiction accuracy, update correctness, and downstream accuracy rise across ρt quintiles, while false-CORRECT falls from 2.1% to 0.6%.ρt tracks retrieval evidence precision, but recall is non-monotone and peaks in middle quintiles.Source: §C.8 · p. 37, Table 24 · p. 38

Correction delay during drift

  • On AMemGym, 14.2% of genuine transitions experience a low-ρt correction delay; 78.0% clear within one period and 94.0% within two.On DiMoBench and DiMoBench-XL, 7.3% and 14.8% of assessable shifts are delayed, respectively; most delays clear within two turns.Source: §C.8 · p. 38

Cold start and genuinely new topics

  • For new-topic cold starts, a first cue is written 96.5% of the time on an empty store, while only 5.6% of distractor cues become hypotheses.With a populated store, a usable belief forms within one period in 84.2% of cases and within two periods in 92.7%.Source: §C.8 · p. 38

C.9 Retrieval and end-to-end latency analysis

PreCoMem reports competitive retrieval and end-to-end latency on LoCoMo, with a controlled latency tail. Its vectorized scoring and deterministic flow are presented as ways to keep memory-system overhead modest.

C.9 Retrieval and end-to-end latency analysis

  • Table 26 reports that PreCoMem achieves retrieval P50/P95 latency of 1.95s/3.86s and end-to-end P50/P95 latency of 2.35s/4.36s.Source: §C.9 · p. 39
  • Confidence-modulated hybrid scoring aggregates semantic and lexical signals through parallelizable matrix operations, avoiding generative reranking or iterative graph traversal.The paper says this keeps overhead modest relative to pure vector retrieval as the belief store grows.Source: §C.9 · p. 39
  • PreCoMem’s retrieval P95 of 3.86s and E2E P95 of 4.36s contrast with A-Mem’s E2E P95 up to 7.68s.The paper attributes PreCoMem’s controlled tail to its deterministic, nonrecursive retrieval flow.Source: §C.9 · p. 39
  • E2E latency exceeds retrieval latency by 0.4s at P50 and 0.5s at P95, with the gap dominated by generation.The paper reports these overheads as comparable to other nonagentic baselines and says PROFILE/CORRECT does not materially inflate generation cost.Source: §C.9 · p. 39

C.10 Structural comparison: why PreCoMem outperforms existing memory systems

The structural comparison argues that PreCoMem couples two properties missing from the compared baselines: confidence-weighted commitment and coherent retrieval. The shared confidence weight links what enters memory with what surfaces at read time.

C.10 Structural comparison: why PreCoMem outperforms existing memory systems

  • The comparison presents these two structural properties as distinct from the nine baselines and alongside the paper’s headline accuracy gains.The passage frames noise robustness and faithful trait preservation, plus retrieval coherence and reduced fragmentation, as the axes governing long-horizon behavior.Source: §C.10 · pp. 39–40
  • Continuous confidence weights and a HYPOTHESIS waiting room defer commitment on weak cues, while soft decay reduces contradicted beliefs without erasing established traits.The passage contrasts this with categorical updates, hard deletion, or overwriting in the compared systems.Source: §C.10 · p. 40
  • Confidence-weighted tri-signal relevance ranks supported items higher and deprecated items lower, while entropy-based reliability damps updates when retrieved score mass is diffuse.The passage describes this as an implicit reranker without an extra LLM call, and connects it to bounded retrieval contamination and effective memory on DiMoBench-XL.Source: §C.10 · p. 40
  • PreCoMem couples commitment and retrieval through one scalar confidence weight, linking a store faithful to core traits with a coherent, currently supported read-time view.The passage contrasts this coupling with baselines that optimize one axis at a time and identifies it as the structural reason for improvements across four main benchmarks.Source: §C.10 · pp. 40–41

C.11 System inspectability, traces, and qualitative analyses

PreCoMem exposes belief contents and provenance for inspection while using the same retrieved evidence to answer and update memory. Traces and qualitative cases illustrate provisional updates, soft correction, and recommendations calibrated to accumulated support; the released prompts document the system’s LLM sub-agents and evaluations.

C.11 System inspectability, traces, and qualitative analyses

  • Belief snapshots expose propositions, type, confidence, timestamps, and episodic provenance for auditing, while confidence metadata stays internal to retrieval rather than the response-generation prompt.Snapshots support inspection, regression testing, and qualitative error analysis without rerunning the model.Source: §C.11 · p. 41
  • A single ambiguous social change cue triggers PROFILE: PreCoMem adds a low-confidence hypothesis instead of immediately correcting Mike’s established social-enthusiast belief.The resulting answer reflects a possible decline, and later repeated cues can reinforce and promote the hypothesis.Source: §C.11 · p. 42
  • After repeated evidence of changed preferences, a strong contradiction triggers CORRECT, softly decaying old social beliefs and helping the answer resist a temporal decoy.The trace reports that recent solitude evidence is retrieved while the older statement is downweighted by correction and recency.Source: §C.11 · p. 43
  • In skill calibration, behavioral cues and teacher feedback temper the credential-based overestimate, yielding fluency practice rather than premature advanced recommendations.The case preserves the belief that the user’s foundation is adequate while lowering confidence in an overly optimistic self-evaluation.Source: §C.11 · p. 43, Figure 11 · p. 43
  • Project reasoning separates established Python/Flask skills from React learning and deferred microservices, producing recommendations matched to demonstrated capability.The case marks React as “loading” and identifies microservices as a challenge zone for later progression.Source: §C.11 · pp. 43–44, Figure 11 · p. 43
  • A new guitar topic enters as a retrieval-visible, low-confidence hypothesis and becomes a FACT only if repeated support raises it past the promotion threshold.Without recurrence, it remains low confidence and does not outrank committed facts.Source: §C.11 · p. 44
  • Evidence used for an answer also updates confidence through MAINTAIN, PROFILE, or CORRECT, connecting response quality to later retrieval and recommendations.The loop incorporates accumulated support, implicit verification, and soft correction under contradiction.Source: §C.11 · p. 44

D Prompts

  • The released prompts include the three operator sub-agents, response generation, benchmark QA and judge prompts, and DiMoBench generation and verification prompts.The paper identifies their source files in the released code.Source: D Prompts · p. 44
Loading 2610.29467v1…