PaperlayerView exact revision on arXiv

Source-linked AI summary

AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents

Shannan Yan, Jingchen Ni, Leqi Zheng, Jiajun Zhang, Peixi Wu, Dacheng Yin, Jing Lyu, Chun Yuan, Fengyun Rao

arXiv:2603.16496v2cs.CL

TL;DR

Long-horizon dialogue agents need memory systems that recover user-relevant evidence beyond semantic similarity, preserve temporal and causal coherence, and adapt memory structure to each question. AdaMem addresses this with heterogeneous user-centric memories, question-conditioned retrieval, and role-specialized evidence processing, achieving state-of-the-art results on LoCoMo and PERSONAMEM. The framework improves answer quality within this scope but increases complexity, token cost, and latency and remains dependent on upstream parsing and backbone reasoning.

  • Problem

    Existing memory systems may miss user-relevant evidence through semantic-only retrieval, weaken temporal and causal coherence through fragmented storage, and use static granularities that do not fit different questions.

  • Method

    AdaMem combines working, episodic, persona, and graph memories with question-conditioned retrieval, target-participant resolution, and role-specialized evidence synthesis and answer generation.

  • Results

    AdaMem achieves state-of-the-art performance on the LoCoMo and PERSONAMEM benchmarks.

  • Takeaways & Limitations

    The results support adaptive memory organization and retrieval for complex multi-session interactions and long-term reasoning scenarios.

  • Takeaways & Limitations

    AdaMem increases system complexity, token cost, and latency while depending on upstream parsing and backbone reasoning.

Abstract

from arXiv · show

Large language model (LLM) agents increasingly rely on external memory to support long-horizon interaction, personalized assistance, and multi-step reasoning. However, existing memory systems still face three core challenges: they often rely too heavily on semantic similarity, which can miss evidence crucial for user-centric understanding; they frequently store related experiences as isolated fragments, weakening temporal and causal coherence; and they typically use static memory granularities that do not adapt well to the requirements of different questions. We propose AdaMem, an adaptive user-centric memory framework for long-horizon dialogue agents. AdaMem organizes dialogue history into working, episodic, persona, and graph memories, enabling the system to preserve recent context, structured long-term experiences, stable user traits, and relation-aware connections within a unified framework. At inference time, AdaMem first resolves the target participant, then builds a question-conditioned retrieval route that combines semantic retrieval with relation-aware graph expansion only when needed, and finally produces the answer through a role-specialized pipeline for evidence synthesis and response generation. We evaluate AdaMem on the LoCoMo and PERSONAMEM benchmarks for long-horizon reasoning and user modeling. Experimental results show that AdaMem achieves state-of-the-art performance on both benchmarks. The code will be released upon acceptance.

1 Introduction

AdaMem addresses three limitations of long-horizon dialogue memory—semantic-only retrieval, fragmented experiences, and static granularity—with adaptive, user-centric structured memories and question-conditioned retrieval. Its role-specialized pipeline is evaluated on LoCoMo and PERSONAMEM, achieving state-of-the-art performance on both benchmarks.

  • Motivation: Long-horizon dialogue requires memory that remains queryable, coherent, and robust as user goals and conversations evolve.Without such organization, memory can become redundant, fragmented, or misaligned with downstream reasoning.
  • Limitations: Semantic retrieval may miss stable preferences, personal attributes, and broader behavioral patterns crucial for user-centric understanding.
  • Limitations: Storing related experiences as isolated fragments weakens temporal and causal coherence during reconstruction and reasoning.
  • Limitations: Fixed-length or coarse memory segmentation can add irrelevant context, while overly fine-grained fragments can obscure dependencies across events and topics.
  • AdaMem: AdaMem organizes dialogue history into working, episodic, persona, and graph-based memories within a unified user-centric framework.The design preserves multiple abstraction levels and targets evidence around the relevant participant.
  • AdaMem: AdaMem combines participant-aware, question-conditioned retrieval with role-specialized evidence synthesis and answer generation.The pipeline resolves target participants, uses relation-aware graph expansion when needed, and separates memory maintenance from answer-time reasoning.
  • Evaluation: AdaMem achieves state-of-the-art performance on the LoCoMo and PERSONAMEM benchmarks.The benchmarks assess long-horizon reasoning and user modeling, respectively.

2 Related Works

Research on agent memory has progressed from context chunking toward structured, modular systems, but user-centric adaptive organization and retrieval remain insufficiently addressed. AdaMem builds on role specialization while targeting consistency and adaptive retrieval over heterogeneous long-horizon dialogue memories.

  • Memory Systems: Earlier memory approaches commonly partition long contexts into smaller chunks, while later systems introduce paging, segmentation, and more modular memory designs.MemGPT is cited as managing long-term memory through paging and segmentation, and Mem0 as an in-memory abstraction.
  • Multi-Agent and Agentic Memory: Multi-agent research emphasizes role specialization, collaborative problem solving, and interactive decision making, while agentic-memory work targets long-term information retention.
  • Research Gap: Existing efforts generally do not explain how long-horizon dialogue evidence should be organized and adaptively retrieved in a user-centric manner.MIRIX introduces specialized memory-organization agents but lacks explicit mechanisms for long-term memory consistency.

3 Approach

AdaMem organizes participant-specific dialogue history across complementary memory structures and uses a target-aware, question-conditioned pipeline to retrieve and synthesize evidence. Its adaptive retrieval combines semantic and relation-aware mechanisms according to question requirements, followed by role-specialized processing for answer generation.

  • Memory construction: AdaMem maintains working, episodic, persona, and graph memories for each participant.These structures preserve recent context, structured long-term records, stable traits, and relation-aware connections.
  • Memory construction: Incoming utterances are normalized into canonical records before updating the four memory structures.The normalized representation includes summaries, topics, factual snippets, attributes, timestamps, and speaker identity.
  • Memory construction: When working memory reaches capacity, AdaMem consolidates the oldest contiguous messages into episodic event, fact, and attribute stores.Independent routers classify evidence as ADD, UPDATE, or IGNORE, while original messages remain available as provenance.
  • Memory construction: Topic regrouping and graph synchronization convert fine-grained records into reusable topic, persona, and relation-aware memory structures.Grouped episodic and attribute records support topic- or aspect-level summaries, while message and consolidated records are indexed in the graph.
  • Question-conditioned retrieval: Before retrieval, AdaMem resolves the target participant and builds a question-conditioned route that selects semantic retrieval, graph expansion, or both.Simple factoid questions generally use lightweight semantic retrieval, whereas temporal or causal questions trigger broader structural exploration.
  • Evidence synthesis and response generation: Retrieved evidence is fused from participant-aware baseline sources and bounded graph expansion, then processed by specialized research and working agents.The Research Agent gathers and integrates evidence, while the Working Agent turns the resulting summary into the final answer.

4 Experiments

AdaMem is evaluated across long-context reasoning and user-modeling benchmarks, multiple model backbones, component ablations, and efficiency settings. Results show strong benchmark performance and complementary contributions from structured memory, evidence fusion, and role-specialized reasoning, with moderate computation overhead.

  • Evaluation Setup: AdaMem evaluates long-context reasoning on LoCoMo and user-representation generalization on PERSONAMEM, using standard benchmark metrics and multiple model settings.LoCoMo reports F1 and BLEU-1, while PERSONAMEM uses accuracy; experiments include closed-source and open-source backbones.
  • Benchmark Results: 44.65% overall F1 on LoCoMo with GPT-4.1-mini represents a +4.4% relative improvement over the previous state-of-the-art method.The largest category gain occurs in temporal questions, reaching +23.4% F1.
  • Benchmark Results: 63.25% accuracy on PERSONAMEM outperforms all baselines by 5.9% relatively, including a 27.3% gain on generalization to new scenarios.The reported gains span a benchmark with evolving user attributes and preferences across multi-session dialogues.
  • Ablation Study: Removing graph memory lowers overall F1 from 44.65 to 42.63, while removing fusion yields 42.77 and replacing multi-agent reasoning yields 43.24 F1.These ablations indicate that graph memory, evidence fusion, and role specialization contribute complementarily.
  • Hyperparameter Sensitivity: K = 10 and Li = 2 provide the default retrieval and deliberation settings, balancing evidence coverage against marginal gains, noise, redundancy, and latency.Increasing K beyond 10 gives marginal gains, while one iteration is insufficient and three slightly reduces performance.
  • Efficiency Analysis: 44.65 F1 is achieved with higher quality but greater computational cost than Mem0, which uses fewer input tokens and lower latency.AdaMem’s additional overhead comes from route planning, graph expansion, and the role-specialized response loop.

5 Conclusion

AdaMem is an adaptive, user-centric memory framework that combines participant-specific memory structures with question-conditioned retrieval and evidence fusion for long-horizon dialogue. Experiments demonstrate its effectiveness, including state-of-the-art performance on LoCoMo.

  • AdaMem combines participant-specific working, episodic, persona, and graph memories with question-conditioned retrieval planning and unified evidence fusion.The framework is designed to retrieve and integrate evidence in a structured, target-aware manner.
  • Experiments on long-horizon reasoning and user modeling benchmarks demonstrate AdaMem's effectiveness.
  • AdaMem achieves state-of-the-art performance on LoCoMo, supporting adaptive memory organization and retrieval for complex multi-session interactions.
  • The results suggest moving beyond uniform storage and fixed retrieval heuristics toward adaptive, question-aware, user-centric memory designs.

Limitations

AdaMem's structured and adaptive design improves answer quality but introduces practical costs and dependencies. Its complexity, resource demands, and reliance on upstream processing constrain deployment and error recovery.

  • AdaMem's structured memories, adaptive retrieval, and role-specialized evidence synthesis increase system complexity, token cost, and latency.
  • The framework depends on upstream parsing and backbone reasoning, making target resolution, entity linking, and temporal normalization errors difficult to recover.
  • The appendix provides implementation details, graph construction details, case studies, and prompt templates in Sections B through E.

B More Implementation Details

The implementation fixes benchmark decision rules in advance and uses a lightweight target resolver that preserves ambiguity rather than forcing premature participant assignment.

  • Benchmark decision rules are fixed before evaluation and reused across runs to attribute gains to retrieval design rather than hidden prompt engineering or benchmark-specific tuning.
  • AdaMem resolves targets as user, assistant, both, or ambiguous based on explicit name mentions.
  • For both or ambiguous targets, AdaMem retrieves from both participant bundles and merges candidates before final reranking instead of forcing early commitment.
  • Owner-aware graph filtering is applied only for unambiguous targets, and no separately trained target-participant classifier is used.

B.2 Route Planning and Graph Expansion

AdaMem plans retrieval from question cues, using graph expansion for temporal or relational questions while keeping suitable single-hop questions on semantic retrieval. Its implementation also uses conservative refinement and sparse, threshold-free clustering.

  • Route Planning and Graph Expansion: The route planner detects temporal, relation, attribute, and single-hop cues before selecting a retrieval strategy.
  • Route Planning and Graph Expansion: Temporal or relation cues enable graph retrieval, while suitable single-hop questions remain on semantic retrieval whenever possible.
  • Route Planning and Graph Expansion: An LLM refines the rule-based plan only below a confidence threshold of 0.75, with clipped outputs for conservative control.
  • Route Planning and Graph Expansion: Each participant's working-memory queue holds 20 messages and consolidates the oldest segment of 5 messages when full.
  • Route Planning and Graph Expansion: AdaMem forms sparse, threshold-free merge clusters by retaining each key's most similar neighbor and taking connected components.

B.5 Graph Propagation Priors

AdaMem assigns fixed strengths to typed graph edges during indexing and uses retrieval-time priors for bounded propagation. Question-aware planning can modestly adjust selected priors, while multi-hop expansion applies a fixed decay.

  • Graph edge priors: Indexing assigns distinct write-time strengths to mention, support, temporal, and speaker-related edges.The supplied passages list fixed values including 0.75 for message–topic mentions, 0.85 for message–fact supports, and 0.65 for speaker_related edges.
  • Retrieval-time propagation: Retrieval uses default edge priors from Table 6 for graph propagation.The table is identified as the source of default retrieval-time edge priors, while the accompanying text describes propagation using these priors.
  • Question-conditioned adjustment: The planner can increase temporal priors for temporal questions and speaker-related priors for attribute-heavy questions.These adjustments are modest and question-dependent rather than freely learned during retrieval.
  • Propagation decay: Multi-hop propagation uses a fixed hop-decay factor λ = 0.85.The decay applies across graph hops during retrieval.

B.6 Fusion Score Components and Weights

AdaMem combines rank- and heuristic-based fusion terms rather than learned scorers. Its benchmark fusion values are fixed before evaluation, with an optional route refiner allowed only clipped per-question adjustments.

  • Fusion components: The four fusion terms in Eq. (3) use simple rank- and heuristic-based priors rather than learned scorers.The recency term decreases linearly over the merged candidate list, while the factual term gives a high bonus to baseline fact-retrieval matches and a small residual bonus otherwise.
  • Benchmark weights: The benchmark configuration uses a default fusion prior shown in Table 7.The values are described as the released benchmark configuration defaults.
  • Weight fitting: Fusion values are fixed before evaluation rather than fit from supervision.An optional route refiner may adjust them per question, but only within clipped ranges.
  • Weight adjustment bounds: The route refiner clips adjustments to α ∈[0.90, 1.00], β ∈[0, 0.08], γ ∈[0, 0.03], and δ ∈[0, 0.03].These bounds constrain question-specific changes to the fusion prior.

B.7 Representative Benchmark Defaults

AdaMem’s released retrieval design builds graph memory from normalized utterance records and typed deterministic edges. At query time, semantic seeds undergo bounded, target-aware multi-hop expansion using lightweight type-specific weights.

  • Graph construction: The graph is built from normalized utterance records using deterministic node typing and edge-construction rules.Graph extraction is not performed by a standalone LLM graph-extraction prompt.
  • Node types: Each utterance creates a message node plus topic, fact, attribute, and event nodes from normalized fields.Message nodes store raw text, speaker, timestamp, and turn index; event nodes are linked from facts or persona snapshots.
  • Edge types: Typed edges encode mentions, support relations, topical continuity, temporal order, and speaker continuity.The listed edge types include mentions, supports, same_topic, temporal_next, and speaker_related.
  • Persona indexing: Persona snapshots are indexed as fact or attribute nodes so long-term user descriptors remain reachable through graph retrieval.This connects persistent user descriptors to the same typed retrieval structure.
  • Query-time retrieval: At query time, AdaMem selects semantic seed nodes and performs bounded multi-hop expansion filtered by the inferred target referent.User-focused questions remain centered on user-grounded evidence, while ambiguous questions can access both participant bundles.
  • Propagation mechanism: Graph propagation uses lightweight type-specific edge weights rather than learned graph reasoning.The graph is therefore presented as a structured retrieval mechanism.

D Case Studies

The case studies show AdaMem succeeding when target identity, personal context, and relations guide retrieval, while revealing failures in implicit entity and temporal alignment. The appendix describes prompts that normalize, update, organize, refine, and iteratively integrate memory evidence.

  • Success case: AdaMem’s success case requires identifying Caroline, isolating a father-related horse activity, and mapping “what activity” to horseback riding.The case is not a pure surface-form lookup despite the answer appearing explicitly in the dialogue.
  • Success case: Target resolution and relation-aware aggregation recover the supporting utterance from Caroline’s memory bundle rather than only a vague semantic neighbor.The case highlights preserving who said what, personal context, and how the event should be reconstructed.
  • Failure case: AdaMem’s failure case requires linking a book title to “this book” and converting “last year” into the absolute year 2022.Missing either implicit alignment leaves retrieval incomplete and the final answer ungrounded.
  • Failure case: The current pipeline relies mainly on retrieval planning, graph expansion, and lightweight temporal salience rather than write-time canonicalization of relative time expressions.The failure exposes temporal grounding and implicit entity linking as unresolved challenges.
  • Message understanding: The message-understanding prompt converts each utterance into a normalized record containing topics, attitude, reason, facts, attributes, summary, and rationale.The resulting record is written into working memory and used by downstream memory-update modules.
  • Memory updating: Router prompts update episodic memory by choosing UPDATE, ADD, or IGNORE for new message-derived items.They compare new items with existing topics, facts, and attributes to handle revisions, additions, or redundancy.
  • Route refinement: The optional route refiner conservatively adjusts graph usage, hop depth, seed count, and fusion weights to avoid unnecessary expansion.Its output includes controls such as use_graph, graph_topn, hop_k, and fusion parameters.
  • Answer pipeline: The Research Agent follows a Planning → Search → Integrate → Reflection loop, while the Working Agent converts the consolidated summary into the final answer.This separates evidence collection from answer realization while retaining a unified memory interface.
Loading 2603.16496v2…