EvoWiki: Incremental State Overwriting and Traceable Question Answering for Cross-Meeting Knowledge Evolution

Dongsheng Chen; Tianyu Wang; Wenhui Que

cs.CLAug 24, 2026

EvoWiki: Incremental State Overwriting and Traceable Question Answering for Cross-Meeting Knowledge Evolution

Authors: Dongsheng Chen, Tianyu Wang, Wenhui Que

Organizations: WeChat, Tencent Inc., Beijing, China

Abstract

In long-term collaboration spanning multiple meetings, factual states such as decisions and risks are continually revised, overturned, and replaced. Existing long-context methods typically stack the entire history, while many RAG and structured-memory methods organize knowledge as static or append-only facts and rely on semantic relevance at read time. Without explicit modeling of knowledge lifecycles, these approaches may retain conflicting old and new states simultaneously or discard history, leading to stale retrieval and answers that are difficult to verify. We present EvoWiki, an incremental question-answering architecture for dynamic long-form text. EvoWiki decouples offline incremental construction (BUILD) from online structured reading (READ). BUILD captures the intra-meeting micro-evolution from proposal to decision and uses entity version chains and a fine-grained State-Overwrite Protocol to explicitly distinguish current valid states from superseded history while preserving meeting-level provenance anchors. READ bypasses relevance-based Top-k retrieval and performs deterministic entity addressing, temporal resolution, and cross-entity multi-hop aggregation over the complete Wiki to produce grounded and traceable answers. We further introduce CrossMeet, a high-fidelity bilingual benchmark designed to simulate long-term state evolution, covering factual consistency, temporal reasoning, and cross-meeting multi-hop reasoning. Across six datasets and two reader models, EvoWiki improves macro-average Judge Accuracy over the strongest baselines by 9.72 and 10.00 percentage points, respectively. Human evaluation shows that EvoWiki is more robust and factually faithful under frequent state flips, validating valid-state-oriented reading as a reliable approach to cross-meeting knowledge evolution.

Figures & tables

Figure 1: EvoWiki architecture. Build writes meetings chronologically into an entity-centered core with version and provenance information. Wiki-only Read uses deterministic addressing and cross-entity temporal aggregation to produce a grounded answer while retaining state-level provenance for evidence-trail recovery.

Metric	CM-EN	CM-ZH
Topics	100	100
Meetings	500	500
QA Pairs	2,000	2,000
Context Tokens (Avg.)	26,861	29,480
Answer Tokens (Avg.)	64.6	108.1
Reasoning Hops (Avg. / Max.)	2.84 / 5	2.77 / 5

Table 1: Statistics of CrossMeet-EN and CrossMeet-ZH. Taxonomy

\kappa

measures question-type classification agreement among three validation models.

Dimension	Type	CM-EN	CM-ZH
Question	Factual Consistency	1000 (50.0%)	1000 (50.0%)
	Timeline Reasoning	600 (30.0%)	600 (30.0%)
	Multi-hop Reasoning	400 (20.0%)	400 (20.0%)
Hops	2 Hops	941 (47.0%)	1005 (50.2%)
	3 Hops	635 (31.8%)	608 (30.4%)
	4 Hops	235 (11.8%)	222 (11.1%)

Table 2: Distribution of question types and reasoning hops in CrossMeet.

Evaluation Dimension	Pass Rate	Fleiss’ $\kappa$
Answer–Evidence	98.3%	0.89
Meeting/Update Coherence	95.7%	0.84
Type/Hop Accuracy	97.7%	0.91

Table 3: Human validation on 300 CrossMeet instances sampled equally from English and Chinese.

Model	Method	CM-EN	CM-ZH	ELITR	LongMemEval	MeetingQA	MuSiQue	Avg
DeepSeek-V4-Flash	Direct LLM	65.90	66.03	59.62	19.80	37.82	49.03	49.70
	VanillaRAG (BM25)	43.98	54.93	48.46	37.60	37.40	45.92	44.72
	VanillaRAG (Dense)	51.10	59.43	55.38	34.60	37.12	46.13	47.29
	VanillaRAG (Hybrid)	51.60	58.75	51.15	39.20	37.06	46.03	47.30
	RAPTOR	49.20	55.20	46.54	40.10	33.84	48.57	45.57
	GraphRAG	56.33	61.30	58.08	39.10	35.40	51.76	50.33

Table 4: Judge Accuracy (%) across six datasets. Best results within each reader block are shown in bold.

Variant	CM-EN	CM-ZH	ELITR	LongMemEval	MeetingQA	MuSiQue	Avg
w/o Overwrite	62.63	65.13	41.54	47.20	36.37	55.88	51.46
w/o Coref	61.03	66.60	43.46	43.20	35.74	55.52	50.93
w/o Entity	63.88	68.23	47.31	49.00	36.65	57.30	53.73
w/o Tags	60.43	62.78	42.31	41.40	35.50	56.29	49.79
EvoWiki	68.78	69.55	60.77	61.60	39.16	60.67	60.09

Table 5: Component ablation of EvoWiki with DeepSeek-V4-Flash, reporting Judge Accuracy (%) across six datasets.

Method	CM-EN	CM-ZH	ELITR	LongMemEval	MeetingQA	MuSiQue	Macro-Avg
Direct LLM	$65.90\pm 0.74$	$66.03\pm 0.68$	$59.62\pm 0.91$	$19.80\pm 1.22$	$37.82\pm 0.95$	$49.03\pm 0.83$	$49.70\pm 0.89$
VanillaRAG	$51.60\pm 1.10$	$58.75\pm 1.05$	$51.15\pm 0.96$	$39.20\pm 1.18$	$37.06\pm 1.02$	$46.03\pm 0.93$	$47.30\pm 1.04$
GraphRAG	$56.33\pm 0.82$	$61.30\pm 0.75$	$58.08\pm 0.79$	$39.10\pm 1.06$	$35.40\pm 0.88$	$51.76\pm 0.81$	$50.33\pm 0.85$
LogicRAG	$54.20\pm 0.76$	$58.50\pm 0.72$	$54.62\pm 0.83$	$48.50\pm 0.97$	$29.42\pm 1.10$	$56.99\pm 0.78$	$50.37\pm 0.86$
EvoWiki	$\mathbf{68.78\pm 0.52}$	$\mathbf{69.55\pm 0.49}$	$\mathbf{60.77\pm 0.58}$	$\mathbf{61.60\pm 0.71}$	$\mathbf{39.16\pm 0.69}$	$\mathbf{60.67\pm 0.55}$	$\mathbf{60.09\pm 0.59}$

Table 6: Performance stability across five runs under the DeepSeek-V4-Flash reader. Results are Judge Accuracy (%) reported as mean

\pm

standard deviation; VanillaRAG denotes the Hybrid variant.

Figure 2: Error attribution for 212 EvoWiki errors, separated into Build and Read causes.

Method	Latency (s)	Read Tokens	Build Tokens
EvoWiki (Ours)	3.60	17,143	114,016
Direct LLM	6.54	26,374	–
VanillaRAG (Hybrid)	1.01	3,481	–
GraphRAG	1.94	9,452	131,298
LogicRAG	4.68	11,238	–
LightRAG	1.78	7,352	67,263

Table 7: Mean query latency and read/build token usage on CrossMeet-EN. Read costs are averaged per query and build costs per project.

Method	Gemini 3.1 Pro	Claude Opus 4.7	DeepSeek-V4-Pro
EvoWiki	60.50	60.17	59.83
Direct LLM	50.17	50.50	49.83
VanillaRAG	46.83	46.50	46.33
GraphRAG	49.83	50.00	49.50
LogicRAG	50.00	50.33	49.33

Table 8: Judge Accuracy (%) on the 300-question subset. Rows denote evaluated methods and columns denote judge models. Agreement with Gemini 3.1 Pro over 1,500 responses is

\kappa=0.901

for Claude Opus 4.7 and

\kappa=0.887

for DeepSeek-V4-Pro. VanillaRAG denotes the Hybrid variant.

Figure 3: Judge Accuracy as the number of state flips increases on CrossMeet-EN.

Figure 4: Judge Accuracy by the relative position of supporting evidence in LongMemEval.

Source (Summary)	The project’s “primary launch risk” evolved twice across meetings. It was first flagged as an SOC2 certification delay ; in Meeting 4 it was updated to insufficient customer onboarding capacity (only 2 customer support staff covering 3 enterprise accounts); in Meeting 5 this risk was overridden again to the third customer’s contract negotiation possibly slipping past Q4, affecting the ARR milestone . The question asks for the primary launch risk at the delivery stage, i.e., it requires the latest (Meeting 5) state, not any overturned earlier risk.
Question	What was the primary launch risk identified for the Nexus Analytics project at the delivery stage?
Reference	Third customer contract negotiation may slip past Q4, affecting ARR milestone
EvoWiki	Third customer contract negotiation potentially slipping past Q4, affecting ARR milestone.
Direct LLM	Customer onboarding capacity. (Landed on the Meeting-4 risk, one version behind—missed the Meeting-5 override.)
VanillaRAG	SOC2 certification delay. (Landed on an even earlier risk that had already been overturned twice.)
GraphRAG	Customer onboarding capacity, with only 2 customer support staff for 3 enterprise accounts. (Same stale Meeting-4 state as Direct LLM.)

Table 9: Case study of cross-meeting launch-risk evolution and final-state recovery.

Figure 5: Pairwise human evaluation of EvoWiki against representative baselines on faithfulness, completeness, and conciseness. Results are percentages of wins, ties, and losses from EvoWiki’s perspective.

Method	CM-ZH	CM-EN	LongMemEval
Direct LLM	1.0%	0.3%	70.6%
VanillaRAG (BM25)	3.6%	8.6%	54.9%
VanillaRAG (Dense)	2.5%	4.1%	55.7%
VanillaRAG (Hybrid)	2.3%	3.3%	54.5%
RAPTOR	3.7%	10.7%	52.6%
GraphRAG	1.3%	1.8%	51.7%

Table 10: Abstention rates on CrossMeet and LongMemEval.

Long-Term Conversational Memory

Long-term conversational memory helps AI assistants retain, update, and use information from past conversations across sessions.

See all 267 papers

Jul 1, 2026

TRACE: State-Aware Query Processing over Temporal Evidence Graphs for Conversational Data

Conversational data is increasingly used as a persistent source of user state for long-running assistants and AI agents. However, querying this data remains challenging because conversations naturally evolve: plans are revised, preferences change, and later messages frequently supersede or contradict earlier information. Existing long-memory pipelines largely treat memories as independent text or vector objects. This approach often retrieves semantically similar but stale evidence, offering limited support for state-aware reasoning. To address this problem, we present TRACE, a query processing framework over temporal evidence graphs for evolving conversational data. TRACE models conversations as a hierarchical graph spanning events, sessions, and topics, enriched with typed temporal, causal, update, and contradiction relations. Crucially, the framework maintains validity annotations so obsolete facts remain accessible for historical queries but are discounted for current-state answers. At query time, TRACE combines vector-based note retrieval with graph-guided evidence search, generating validity-aware support paths and a hybrid context for answer generation. This design separates lexical recall from evidence reconstruction, enabling bounded query-time reasoning over long conversational histories. Experiments on long-conversation query-answering (QA) benchmarks show that TRACE improves temporal and multi-hop reasoning, with ablations highlighting the importance of hierarchy, update-aware seeding, and path-grounded evidence.

Maolin Wang, Yu Wang, Zichun Liu +5

Hong Kong Institute of AI for Science, City University of Hong Kong, Hong Kong SAR, China · Independent Researcher, Beijing, China · Central South University, Changsha, China +1

Multi-Hop QA Graph Retrieval-Augmented Generation Temporal IR Long-Term Conversational Memory Temporal KGs Temporal QA

Ask PDF

Apr 30, 2026

EviMem: Evidence-Gap-Driven Iterative Retrieval for Long-Term Conversational Memory

Long-term conversational memory requires retrieving evidence scattered across multiple sessions, yet single-pass retrieval fails on temporal and multi-hop questions. Existing iterative methods refine queries via generated content or document-level signals, but none explicitly diagnoses the evidence gap, namely what is missing from the accumulated retrieval set, leaving query refinement untargeted. We present EviMem, combining IRIS (Iterative Retrieval via Insufficiency Signals), a closed-loop framework that detects evidence gaps through sufficiency evaluation, diagnoses what is missing, and drives targeted query refinement, with LaceMem (Layered Architecture for Conversational Evidence Memory), a coarse-to-fine memory hierarchy supporting fine-grained gap diagnosis. On LoCoMo, EviMem improves Judge Accuracy over MIRIX on temporal (73.3% to 81.6%) and multi-hop (65.9% to 85.2%) questions at 4.5x lower latency. Code: https://github.com/AIGeeksGroup/EviMem.

Yuyang Li, Yime He, Zeyu Zhang +1

The Australian National University · UNSW Sydney

Temporal IR Evidence Retrieval Persistent Memory for Language Models Multi-Hop QA Information Retrieval Long-Term Conversational Memory+1

Ask PDF

Aug 2, 2026

TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents

Large language model agents have shown strong capabilities in generating coherent and contextually appropriate responses, yet robust long-horizon dialogue remains limited by the lack of external memory that is traceable, updatable, and diagnostically transparent. Existing memory-augmented agents often store memories as isolated records or overwritable states, making it difficult to preserve how information originates, evolves, conflicts, or becomes obsolete over time. We propose TrajWiki, a trajectory-based memory framework for long-horizon conversational agents. Instead of treating memory as static entries, TrajWiki represents each memory as a source-grounded evolution trajectory, maintained through immutable episodic snapshots and claim-level operations such as ADD, REVISE, and DEPRECATE. To reduce fragmentation and retrieval cost, TrajWiki further introduces Memory Wiki, a persistent intermediate layer that incrementally compiles dialogue history into structured and interlinked wiki pages capturing salient entities, events, quantities, topics, and conflicts. At inference time, queries are routed hierarchically from relevant wiki pages to linked memory trajectories, then to corresponding snapshots and source messages for evidence-grounded answer synthesis. Experiments on LoCoMo and MedMT show that TrajWiki improves long-horizon dialogue performance across both open-source and closed-source LLM backbones, while providing greater interpretability and diagnostic visibility into memory evolution, retrieval failures, and answer generation.

Jingyu Sun, Yuyang Xue, Mingyang Li +9

1The University of Manchester · 2The University of Melbourne · 3The University of Edinburgh +2

Agent Memory LLM Agent Memory Persistent Memory for Language Models Long-Term Conversational Memory Long-Horizon LLM Agents Agent Memory Management

Ask PDF

Abstract

Long-Term Conversational Memory

Explore similar work

TRACE: State-Aware Query Processing over Temporal Evidence Graphs for Conversational Data

EviMem: Evidence-Gap-Driven Iterative Retrieval for Long-Term Conversational Memory

TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents