PaperlayerView exact revision on arXiv

Source-linked AI summary

SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue

Haobo Zheng, Tan Tang, Yan Chen, Weijie Wang, Yingcai Wu

arXiv:2609.26780v2cs.CLcs.AIcs.IRcs.LG

TL;DR

Multi-party memory must recover attribution and changing state from histories where evidence spans participants, groups, and time. SPEAKERMEM-R1 combines speaker-labeled verbatim messages with person- and group-level derived views, then composes query-specific evidence; it exceeds mainstream-framework bests across three benchmarks, including a 69.2% SocialMemBench accuracy, 12.4 points higher.

  • Problem

    Multi-party conversational memory must preserve attribution and relations while reconstructing states from clues distributed across people, groups, and time.

    Source: §1 · p. 1

  • Method

    SPEAKERMEM-R1 pairs speaker-labeled verbatim messages with provenance-linked person- and group-level derived states, then organizes query evidence by participant, event, and time.

    Source: §1 · p. 2, §3 · p. 3, §3.2 · p. 5

  • Results

    SPEAKERMEM-R1 exceeds mainstream-framework bests across GroupMemBench, SocialMemBench, and EverMemBench; its 69.2% SocialMemBench accuracy is 12.4 points higher.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Ablations show that the verbatim and structured tracks, and the person-level and group-level views, contribute complementarily under the standardized evaluation interface.

    Source: §6 · p. 9

    The system assumes reliable rosters, source/owner attribution, and temporal identification; aliases, membership changes, implicit audiences, and parallel events remain challenging.

    Source: §6 · p. 9

Abstract

from arXiv · show

Long-term conversational memory in multi-party settings requires more than retrieving relevant content from long-term conversations: it must distinguish who said what, whom each statement concerns, how individuals perceive one another, what information is shared by the group, and how states change over time. Recent studies on multi-party dialogue benchmarks show that existing general-purpose LLM memory systems tend to lose person and group relations or struggle to integrate clues distributed across members, groups, and time. Together, these issues reveal two core bottlenecks: message attribution and relational understanding in multi-party dialogue, and state reconstruction from interleaved histories. To address both, we propose $\textbf{SpeakerMem-R1}$: its dual-track memory stores speaker-labeled verbatim messages and derived states organized into person-level and group-level views, then combines evidence from both tracks by entity, event, and time at query time. To reduce attribution and update errors during structured memory construction while enabling local deployment, we train Writer-R1 with SpeakerLevenshtein and speaker-conditioned GRPO. On GroupMemBench, SocialMemBench, and EverMemBench, SpeakerMem-R1 achieves binary accuracies of 47.9%, 69.2%, and 61.9%, respectively. On the publicly reported EverMemBench leaderboard from EverMind-AI, we achieves 62.33%, the best reported result among the latest state-of-the-art frameworks. It also achieves 70.85% on all 1,986 LoCoMo questions, which we use as a two-person long-term conversation boundary test. In a controlled evaluation of 305 questions, RL raises the SFT Writer's mean accuracy from 57.38% to 68.20%. We report both binary accuracy and token-F1, and ablations show that the verbatim and structured tracks, as well as person-level and group-level views, are complementary under the standardized evaluation interface.

1 INTRODUCTION

Multi-party memory must preserve who said what, personal versus shared scope, and evolving relations across interleaved histories. SPEAKERMEM-R1 combines speaker-labeled messages with structured state views and reports gains across three benchmarks.

1 INTRODUCTION

  • Existing memory work mainly targets single-user or two-person histories, while group chats add speaker relations, reply structure, cross-topic branches, and state revisions.The paper describes GroupMemBench, SocialMemBench, and EverMemBench as showing degradation by general-purpose systems in multi-party or long-term settings.Source: §1 · p. 1
  • Multi-party memory must address message attribution and state reconstruction across members, groups, and time, not merely retrieve relevant text.Global retrieval may miss low-frequency members or the correct branch, while summaries and generic graphs may lose source, owner, scope, or state versions.Source: §1 · p. 2
  • SPEAKERMEM-R1 preserves speaker-labeled messages alongside provenance-linked derived states organized into person-level and group-level views.At query time, Anchor–Separate–Resolve–Compose organizes evidence by participant, event, and time.Source: §1 · p. 2
  • The authors train a locally deployable Writer with SpeakerLevenshtein and speaker-conditioned GRPO to reduce attribution and update errors while freezing query and answer modules.The contribution summary reports that the RL Writer reaches 95.4% of the LLM Writer reference accuracy on 305 controlled questions.Source: §1 · p. 2
  • 47.9%, 69.2%, and 61.9% binary accuracy on GroupMemBench, SocialMemBench, and EverMemBench exceed mainstream-framework bests by 3.3, 12.4, and 9.4 points, respectively.In the controlled 305-question evaluation, Writer-R1 reaches 68.20%, 10.82 points above SFT.Source: §1 · p. 2

2 RELATED WORK

Prior work spans long-term memory, retrieval and structured stores, and learned memory management. The paper distinguishes its focus as preserving multi-party attribution and training owner-level writing while retrieval and answering remain fixed.

2 RELATED WORK

  • Multi-party benchmarks evaluate group dynamics, social relations and norms, and cross-group collaboration with evolving states; their evidence is distributed across members, groups, and time.The cited benchmarks include GroupMemBench, SocialMemBench, and EverMemBench (Yang et al., 2026; Owolabi, 2026; Hu et al., 2026b).Source: §2 · p. 2
  • Retrieval and memory-graph methods support matching, compression, association, and temporal organization but do not necessarily preserve multi-party attribution, cross-person cognition, or group consensus.The related-work discussion covers lexical and semantic retrieval, persistent-memory systems, and graph-based approaches.Source: §2 · pp. 2–3
  • Learnable-memory work explores memory actions, hierarchy, management, state-difference rewards, and long-horizon credit assignment; this paper focuses on owner-level writing supervision in group chats.The Writer is trained while retrieval and answering modules remain fixed.Source: §2 · p. 3

3 METHOD

SPEAKERMEM-R1 writes traceable verbatim messages and derived states, then reconstructs query-specific evidence for a frozen answerer. Its query process constrains evidence by participants, issue, scope, and time.

3 METHOD

  • The architecture writes a traceable dual-track memory from multi-party messages, reconstructs query-specific evidence, and passes that evidence to a frozen answerer.The Writer converts local messages into structured memory actions; retrieval and answering operate on the resulting memory.Source: §3 · p. 3, Figure 2 · p. 4

3.1 PROBLEM DEFINITION: FROM RELEVANCE RETRIEVAL TO ATTRIBUTED EVIDENCE

  • Ordinary retrieval optimizes evidence relevance, but multi-party QA also requires compatibility in participants, attribution, scope, event, and time.A derived record distinguishes content, source, owner, PERSON/GROUP scope, event, time, state, and a source reference; source and owner can differ.Source: §3.1 · p. 4

3.2 FIVE-LAYER TRACEABLE DUAL-TRACK MEMORY

  • System 1 stores verbatim messages with speaker, time, and channel; System 2 organizes derived states into PERSON-scoped Core and Profile and GROUP-scoped Interaction and Insight layers.Derived records link back to supporting messages, while verbatim evidence provides exact wording and local context.Source: §3.2 · p. 5

3.3 QUERY-CONDITIONED EVIDENCE

  • Project compiles query constraints, System 1 retrieves wording and context, and System 2 selects records by PERSON/GROUP rows, issue, relation, event, and time.Compose organizes the two evidence paths along persons, relations, and update chains; when a derived row is empty, the system falls back to corresponding verbatim messages.Source: §3.3 · p. 5

4 WRITER TRAINING METHOD

Writer training replaces prompt-based writing with a locally deployable small model while query organization and answering remain frozen. Its rewards combine owner-level state quality with downstream QA gain, optimized through position-aligned group-relative clipped GRPO.

4 WRITER TRAINING METHOD

  • Writer-R1 learns ADD/UPDATE/NOOP decisions while query organization and answering remain frozen.The section frames RL as a way to replace an expensive prompt-based Writer with a locally deployable small model.Source: §4 · p. 5
  • SpeakerLevenshtein matches records one-to-one within owner buckets using token-level F1 and normalized sequence matching; omissions and overwriting are penalized.Owner buckets include personal records, GROUP records, and cross-person observations.Source: §4 · pp. 5–6
  • Eight writing trajectories form group-relative advantages at aligned positions, and clipped GRPO updates only the Writer.The reward weights are 0.20 for validity, 0.45 for memory, and 0.35 for QA; positions without effective within-group variation are omitted.Source: Figure 3 · p. 6, §4 · p. 6
  • The local structural potential combines macro-average state quality with a worst-owner term, weighted 0.80 and 0.20, respectively.The worst-owner term prevents frequent people from masking infrequent people or GROUP.Source: §4 · p. 6
  • The global signal is System 1+System 2 QA minus System 1-only QA on the same evidence, excluding questions already answerable from verbatim memory.It measures the QA gain after adding structured memory.Source: §4 · p. 6

5 EXPERIMENTS

Across multi-party benchmarks, SPEAKERMEM-R1 outperforms the strongest mainstream baselines, while ablations support complementary evidence from both memory tracks and hierarchy levels. Controlled writer results show a gain over SFT, and LoCoMo serves as a two-person boundary test.

5.2 MAIN MULTI-PARTY RESULTS

  • 47.9%, 69.2%, and 61.9% are SPEAKERMEM-R1’s highest accuracies on GroupMem, SocialMem, and EverMem, exceeding each benchmark’s best mainstream baseline by 3.3, 12.4, and 9.4 points.Full context is feasible only on SocialMemBench, where SPEAKERMEM-R1 reaches 69.2% versus 69.4% for Full context.Source: Table 1 · p. 7, §5.2 · p. 7

5.3 DUAL-TRACK AND HIERARCHICAL ABLATIONS

  • Removing either memory track reduces accuracy on all three benchmarks, and keeping only person-level or group-level records also degrades performance.Figure 4 defines the track and hierarchy removals; the complementary-evidence comparison is against the better full-system result with or without ASK.Source: Figure 4 · p. 8, §5.3 · p. 8

5.2 MAIN MULTI-PARTY RESULTS

  • 62.33% is SPEAKERMEM-R1’s accuracy on the public EverMemBench leaderboard, ahead of approximately 60.08% for EverOS and 54.75% for RippleMem.The comparison uses all 2,400 questions; public totals are reconstructed from rounded category results, while SPEAKERMEM-R1 uses question-level records.Source: §5.2 · p. 7, Table 2 · p. 8
  • Strong Single, Const, Proact, and Update results contrast with weaker Multi, Skill, and Role scores, which expose cross-evidence, preference, and role-attribution bottlenecks.Source: §5.2 · p. 8
  • On EverMem open-ended questions, SPEAKERMEM-R1 achieves 40.0% versus BM25’s 27.0%, while binary accuracy and token-F1 are not monotonic across variants.The full system has lower SocialMem accuracy than † but higher token-F1 (32.7 versus 27.4); extra people or incorrect scope can invalidate an overlapping answer.Source: §5.2 · p. 8

5.3 DUAL-TRACK AND HIERARCHICAL ABLATIONS

  • With ASK disabled, accuracies are 47.0/69.2/60.5% on GroupMem/SocialMem/EverMem, versus 47.9/64.9/61.9% with ASK.Supplementary retrieval recovers dispersed clues on GroupMem and EverMem but can add redundant evidence on SocialMem.Source: §5.3 · p. 8

5.4 DOES THE FULL R1 OBJECTIVE IMPROVE WRITING?

  • 68.20±0.66% versus SFT’s 57.38±0.33% is Qwen2.5-3B Writer-R1’s accuracy on 305 held-out SocialMem questions, a gain of 10.82 points.The LLM writer reference reaches 71.48±0.66%; the result is not evidence of broad cross-domain RL generalization.Source: §5.4 · pp. 8–9, Table 3 · p. 9
  • 70.85% (1,407/1,986) is SPEAKERMEM-R1’s accuracy on two-person LoCoMo, where multi-hop and open-domain questions remain weak.The paper treats this as a boundary test, not a replacement for multi-party validation.Source: Table 4 · p. 9, §5.4 · p. 9

6 CONCLUSION

The paper presents dual-track memory as a practical approach to multi-party attribution and state reconstruction, while identifying assumptions and open challenges that constrain the system.

6 CONCLUSION

  • The system combines source-linked verbatim evidence with person-level and group-level structured views to support attribution, scope control, and state reconstruction.Ablations indicate that both tracks and both views contribute complementary benefits.Source: §6 · p. 9
  • In the controlled study, the 30-step RL Writer reaches 95.4% of the 71.48% LLM Writer reference accuracy.The conclusion presents this as evidence that a locally deployable small Writer can approach the large Writer's performance while keeping the pipeline practical.Source: §6 · p. 9
  • The system assumes reliable rosters, source/owner attribution, and temporal identification; aliases, membership changes, implicit audiences, and parallel events remain challenging.Future work targets construction and retrieval cost, cross-evidence reasoning, open-domain QA, and generalization across domains and languages.Source: §6 · p. 9

APPENDIX

  • The appendix supplies data structures, algorithms, implementation and evaluation protocols, training details, complete category results, and prompts.It also includes a LoCoMo breakdown and retrieval-budget sensitivity.Source: APPENDIX · p. 15

A IMPLEMENTATION DETAILS

The implementation keeps verbatim messages and derived records linked through provenance, with updates preserving prior states and query-time retrieval organized by person or group.

A IMPLEMENTATION DETAILS

  • The system stores verbatim messages and the channel roster first, then has the Writer produce derived records with source provenance.System 1 selects speaker-attributed messages; System 2 retrieves structured records by per-speaker or group rows and follows provenance pointers to source text.Source: A IMPLEMENTATION DETAILS · p. 15
  • The main experiments, ablations, and controlled Writer evaluation use fixed budgets: System 1 top-k = 10 with recall-n = 40, and System 2 k = 2 records per owner row with source-k = 1.The two retrieval paths use independent budgets.Source: A IMPLEMENTATION DETAILS · p. 15
  • At query time, Project parses roster, issue/event, temporal mode, and relation constraints before evidence is selected, separated into person/group rows, resolved, and composed.The answerer receives only the final evidence set, while R1 training updates only the Writer.Source: A IMPLEMENTATION DETAILS · p. 15

A.1 FIVE-LAYER MEMORY SCHEMA

  • The five-layer schema records implementation layer names, scope, retained information, and each layer's role in query-time evidence construction.System 2 records include source/owner fields and provenance pointers to System 1 messages.Source: §A.1 · p. 15
  • Updates append new nodes and mark earlier nodes as superseded rather than overwriting history.Head/full query mode selects either the current state or the complete update chain.Source: §A.1 · p. 15

B DATA STRUCTURES AND ALGORITHMIC DETAILS

The appendix details how messages become persistent memory records and how query-time retrieval combines verbatim and structured evidence while preserving provenance.

B.2 ONLINE WRITING

  • Online writing stores each message verbatim, then has the Writer propose ADD, UPDATE, or NOOP actions that are validated before valid records are persisted.For updates, new nodes link to prior states; old nodes are retained, and NOOP leaves the derived track unchanged.Source: §B.2 · pp. 15–17

B.1 MEMORY RECORD STRUCTURE

  • The verbatim and derived tracks share persistent record fields, while the Writer generates action fields and the system fills remaining coordinates from messages and current state.System 1 is the verbatim track; System 2 contains four derived layers.Source: §B.1 · p. 15, Table 6 · p. 16, Table 5 · p. 16

B.2 ONLINE WRITING

  • The Writer decides what to record, while deterministic code fills message coordinates, provenance, and update chains.The appendix describes this as separating what to write from coordinates completed by the system.Source: §B.2 · p. 17

B.3 DUAL-TRACK COMPOSITION AT QUERY TIME

  • Query-time composition retrieves verbatim evidence and structured records by issue and roster row, preserves owner, source, time, and provenance, and sends both tracks to the frozen answerer.The two retrieval paths have independent budgets; structured memory determines objects, scopes, and states to check, while verbatim memory supplies exact wording and details.Source: §B.3 · p. 17

C BENCHMARKS AND EVALUATION PROTOCOL

The evaluation section specifies benchmark scope, common settings, and metric definitions, including distinct question-level and network-level ways of averaging SocialMemBench scores.

C BENCHMARKS AND EVALUATION PROTOCOL

  • Main experiments and ablations use System 1 top-k = 10 and System 2 k = 2/source-k = 1, with full context included only as a SocialMem reference upper bound.For the primary three-benchmark comparison, systems follow their official code, recommended configuration, and prompt, while data, metric definitions, judge, and evaluation interface are standardized.Source: C BENCHMARKS AND EVALUATION PROTOCOL · p. 17
  • The evaluation reports question counts, primary metrics, and result locations for its benchmark categories.The evaluation-scope table summarizes the number of questions, metrics, and where category-level results appear.Source: C BENCHMARKS AND EVALUATION PROTOCOL · p. 17, Table 7 · p. 18
  • Token-F1 uses normalized unique token sets and measures lexical overlap rather than semantic correctness; binary accuracy and token-F1 are reported separately.Token-F1 is computed per question and averaged with equal question weights; an empty or disjoint token-set intersection yields zero.Source: C BENCHMARKS AND EVALUATION PROTOCOL · pp. 17–18
  • SocialMemBench's MeanQ weights every question equally, whereas MeanN first averages within each network and then weights networks equally.Both average item-level partial scores rather than the proportion of questions answered completely correctly; the reported 95% CI applies only to MeanN and bootstraps 43 network means.Source: C BENCHMARKS AND EVALUATION PROTOCOL · p. 18

D REPRODUCTION CHECKLIST AND EXPERIMENTAL BOUNDARIES

The reproduction checklist standardizes selected evaluation elements while retaining benchmark-specific system configurations. It also specifies token-accounting rules and the ingestion settings used for cost comparisons.

D REPRODUCTION CHECKLIST AND EXPERIMENTAL BOUNDARIES

  • Evaluations follow each system’s official code, recommended configuration, and prompt, while standardizing data, metrics, the judge, and interface where specified.Implementations may retain speaker/source metadata where supported; the checklist also distinguishes system-stage language models from question-level judging settings.Source: D REPRODUCTION CHECKLIST AND… · p. 18, Table 8 · p. 19
  • Token totals include ingestion input/output, QA input, and final-answer output, exclude judging, and use benchmark question counts as the per-question denominator.SpeakerMem uses session-level ingestion for SocialMem and Window25 for GroupMem/EverMem; Mem0, A-MEM, and HippoRAG use chunk5 ingestion.Source: D REPRODUCTION CHECKLIST AND… · p. 19, Table 9 · p. 20
  • The three benchmarks contain 1,031 SocialMemBench, 745 GroupMemBench, and 2,400 EverMemBench questions.The breakdowns report 348 SpeakerMem sessions for SocialMemBench, 4,815 Writer calls for GroupMemBench, and 2,049 for EverMemBench.Source: D REPRODUCTION CHECKLIST AND… · p. 19

E ADDITIONAL R1 TRAINING DETAILS

The appendix details Writer-R1’s training signals and data, then documents diagnostic dimensions, benchmark category reporting, and alternate-model evaluation settings. It distinguishes question-level binary accuracy from token-F1 and describes the scope of each evaluation configuration.

E ADDITIONAL R1 TRAINING DETAILS

  • SpeakerLevenshtein matches records using coordinate-gated content similarity, while the update-chain reward separately checks old and new states after owner-and-layer alignment.The old- and new-state weights are both 0.50.Source: E ADDITIONAL R1 TRAINING DETAILS · p. 21
  • Writer training uses 15 complete networks split at network level, yielding 73 writing segments, 89 terminal QA items, and 452 supervised actions.Real-network trajectories come from DeepSeek-V4-Flash as proxy supervision; manually authored networks cover UPDATE/NOOP cases.Source: E ADDITIONAL R1 TRAINING DETAILS · pp. 22–23
  • 68.20 ± 0.66% mean accuracy for the RL Writer versus 57.38 ± 0.33% for SFT, a gain of 10.82 percentage points on 305 questions.The query and answer modules were frozen under the same no-ASK configuration; the RL Writer reached 95.4% of the DeepSeek-V4-Flash Writer’s 71.48 ± 0.66% mean accuracy.Source: E ADDITIONAL R1 TRAINING DETAILS · p. 23

F.3 BINARY ACCURACY WITH DEEPSEEK-V4-FLASH AS THE ANSWER MODEL

F.1 FIVE ANALYSIS DIMENSIONS: DEFINITIONS AND TYPICAL ERRORS

  • Five diagnostic dimensions organize recurring errors: member coverage, source-versus-owner attribution, personal/group scope, event disambiguation, and temporal updates with multi-hop reasoning.They are mapped to dominant failure modes, not independently human-annotated labels or a validated taxonomy.Source: §F.1 · pp. 23–24

F.2 CATEGORY DEFINITIONS FOR THE THREE BENCHMARKS

  • GroupMemBench reports six official categories, SocialMemBench nine, and EverMemBench nine memory behaviors; these remain the quantitative reporting units.The diagnostic dimensions are mapped onto official categories, which may involve more than one dimension.Source: §F.1 · p. 23, §F.2 · pp. 24–25

F.4 CATEGORY-LEVEL TOKEN-F1

  • Token-F1 is reported separately from question-level binary accuracy: category-level results cover GroupMemBench and SocialMemBench, while EverMemBench token-F1 is overall only.The two multi-party benchmarks’ token-F1 values are computed directly against references, not from GPT-4o-mini judgments.Source: §F.4 · p. 26

F.5 GPT-5.6-LUNA MODEL CONFIGURATION

  • The GPT-5.6-luna robustness configuration uses GPT-4o-mini for binary judgments and reports category results for GroupMemBench and SocialMemBench.Full context appears only for SocialMemBench; EverMemBench is reported only overall under this configuration.Source: §F.5 · pp. 26–27

G FULL LOCOMO CATEGORY BREAKDOWN

The LoCoMo breakdown tests SpeakerMem-R1 on two-person long-term conversations across five question categories. Its reported pattern is stronger on single-hop and temporal questions and weaker on multi-hop and open-domain questions.

G FULL LOCOMO CATEGORY BREAKDOWN

  • LoCoMo’s five categories are Single-hop, Adversarial/AD, Temporal, Multi-hop, and Open-domain.They cover one-step recall, misleading questions, temporal relations, multi-fact reasoning, and open-domain or cross-context reasoning, respectively.Source: G FULL LOCOMO CATEGORY BREAKDOWN · p. 27
  • Single-hop and Temporal results indicate effectiveness on two-person long-term conversations, while Multi-hop and Open-domain remain major weaknesses.The paper attributes the weakness pattern to cross-segment composition and external knowledge; LoCoMo contains 1,986 questions across five categories.Source: G FULL LOCOMO CATEGORY BREAKDOWN · p. 27

H RETRIEVAL TOP-k SENSITIVITY

This section examines retrieval top-k choices and documents the associated memory-writing, query-projection, verbatim and supplementary retrieval, answer-generation, and evaluation prompts for multi-party dialogue memory. It also specifies partial-credit rubrics for attribution, theory-of-mind, and temporal-shift answers.

H RETRIEVAL TOP-k SENSITIVITY

  • Increasing S1 top-k from 5 to 10 to 20 improved both metrics at every S2 budget, but average Acc gain per added S1 slot fell across both increments.At S2 k = 4, Acc rose from 56.39% to 64.26% to 73.11%, while token-F1 rose from 23.85% to 25.00% to 26.53%.Source: H RETRIEVAL TOP-k SENSITIVITY · p. 29
  • Increasing S2 k from 2 to 4 improved both metrics at all S1 budgets, while increasing it to 6 could reduce performance, including at S1 top-k = 20.At S1 top-k = 20, S2 k = 4 to 6 reduced Acc from 73.11% to 71.15% and token-F1 from 26.53% to 25.78%; redundancy or distraction is a plausible, not established, explanation.Source: H RETRIEVAL TOP-k SENSITIVITY · p. 29
  • S1 top-k = 20 and S2 k = 4 jointly achieved the best observed Acc (73.11%) and token-F1 (26.53%) across the tested grid.The study tested nine configurations with fixed memory, no ASK, and S2 source-k = 1; the main experiments instead use S1 top-k = 10 and S2 k = 2/source-k = 1.Source: H RETRIEVAL TOP-k SENSITIVITY · p. 29, Figure 6 · p. 30
  • The effect of increasing S2 k from 4 to 6 depended on S1 coverage: both metrics rose at S1 top-k = 5 but fell at S1 top-k = 20.This supports controlling the two retrieval budgets separately rather than maximizing both together.Source: H RETRIEVAL TOP-k SENSITIVITY · p. 29, Figure 6 · p. 30

I.1 PROMPT INVENTORY

  • The prompt inventory identifies each prompt’s inputs, outputs, and frozen boundary, with Writer, query, answer, and judging prompts fixed in the main experiments and ablations except for Writer.The prompt snapshots are presented as copyable specifications of module responsibilities and output formats.Source: I COMPLETE PROMPTS · p. 30, §I.1 · p. 31, Table 24 · p. 31

I.2 WRITE: STRUCTURED WRITING

  • The Writer receives the channel roster, latest derived-memory state, and a new message segment, then generates ADD, UPDATE, or NOOP; the system adds validated metadata afterward.The prompt separates speaker-labeled verbatim messages from derived memory and specifies person-level and group-level derived layers.Source: §I.2 · p. 31

I.3 PROJECT: QUERY PROJECTION

  • Project outputs an issue, rows, and tense to determine the query’s memory slice, while roster and within-row retrieval enforce channel and source/owner constraints.Rows select person-level or group-level views, and tense determines whether to use current nodes alone or expand update chains.Source: §I.3 · p. 34
  • Query projection selects person or GROUP rows and head or full tense according to the question’s target and whether it requires current or historical state.For cross-person questions, retrieve the row for the person being discussed; use ALL when unsure which members matter. Past-time questions require full history because the current value may have changed.Source: §I.3 · pp. 34–35

I.4 SELECT AND EXPAND: VERBATIM-TRACK SELECTION

  • Verbatim selection ranks System 1 candidates by necessity and expands sparingly for missing exact values or event details, not missing per-person coverage.Expansion defaults to empty and is limited to at most two anchors; coverage across people is handled by System 2.Source: §I.4 · p. 36

I.5 SUFFICIENCY AND ASK: SUPPLEMENTARY RETRIEVAL DECISION

  • The sufficiency module checks whether System 1 evidence contains the answer and, if not, triggers one supplementary retrieval and reranking with a terse search term.On-topic but vague material does not count as sufficient; missing only per-person coverage does, because another retrieval route handles it.Source: §I.5 · pp. 36–37

I.6 ANSWER: FROZEN ANSWERER

  • The frozen answerer combines speaker- and time-labeled raw excerpts with person/group memory, preserving attribution and distinguishing current states from relevant history.A (none) cell means no derived record on that matter and answers “never” under the prompt’s rule; raw excerpts remain valid evidence when structured memory is silent.Source: §I.6 · pp. 37–38

I.7 EVALUATION PROMPTS

  • Evaluation uses binary judgments that accept equivalent paraphrases, while SocialMem scoring gives partial credit and treats wrong-person attribution as a critical error capped at 0.3.SocialMem MeanQ averages item scores; MeanN averages within dialogue networks before averaging across networks.Source: §I.7 · pp. 38–39
  • Attribution earns full credit for naming the correct speaker without naming the foil, with lower scores for naming the foil or answering incorrectly.The rubric assigns 0.7 when the foil is named and 0.0–0.3 for a wrong answer.Source: §I.7 · p. 39
  • Theory-of-mind answers require both a preference and the person who revealed it through observable action; temporal-shift scores reward old state, new state, and trigger.For temporal shifts, each component contributes 0.33; both states without a trigger score 0.7, and one state scores 0.4.Source: §I.7 · p. 40
Loading 2609.26780v2…