PaperlayerView exact revision on arXiv

Source-linked AI summary

APDMem: Agent-Controlled Progressive Disclosure for Query-Adaptive Long-Term Memory

Chin-Lun Fu, Anagha Kulkarni, Hong Ni, Behrouz Madahian

arXiv:2610.02472v1cs.CLcs.AI

TL;DR

Personalized assistants must recover scattered evidence across long conversation histories. APDMem organizes memory in layers and reveals finer details as needed, then structures retrieved evidence for grounded answers; it reports strong LongMemEval performance while accessing only 8% of conversations.

  • Problem

    Personalized agents must recover user preferences, facts, temporal sequences, and recurring themes scattered across long-term, multi-session interactions.

    Source: §1 · p. 1

  • Method

    APDMem stores conversations in four progressively detailed layers, discloses finer evidence on demand, and structures retrieved evidence and a draft answer into organized notes.

    Source: §1 · pp. 1–2

  • Results

    8% of conversations accessed accompanies strong LongMemEval performance; APDMem reports 87.8% accuracy with GPT-4.1.

    Source: §1 · p. 2, §5 · p. 7

  • Takeaways & Limitations

    Disclosure depth adapts to question type: most questions use the high-level key-fact layer, while some detail-oriented categories benefit from raw-message access.

    Source: §5 · p. 7

    Retrieval recall depends on L0 thematic-summary quality, because omitted relevant details can lower BM25 and dense-retrieval scores.

    Source: Limitations · p. 7

Abstract

from arXiv · show

Personalized LLM assistants must recover sparse evidence from long conversation histories across queries of varying complexity. We introduce APDMem (Agent-controlled Progressive Disclosure Memory), a hierarchical long-term memory architecture that applies progressive disclosure to memory retrieval. Rather than relying on a flat memory store or fixed retrieval granularity, APDMem represents conversation history as four progressively detailed layers: thematic summaries, personalized key facts, turn-level evidence notes, and raw messages. At inference time, a controller applies progressive disclosure to the memory hierarchy: it first reads high-level summaries and drills into finer evidence only when needed. This creates an adaptive cost-fidelity trade-off: simple queries can terminate early, while complex temporal, multi-hop, or exact-evidence queries trigger deeper inspection. A note synthesizer converts retrieved evidence into a query-focused structure that consolidates facts, orders events, and flags contradictions before final answer generation. Experiments on LongMemEval show that APDMem achieves strong performance for long-context memory reasoning while accessing only 8% of the total conversations.

1 Introduction

Personalized agents must retain and recover user-specific facts across long, multi-session conversations where relevant information may be scattered. APDMem addresses this with layered memory, query-adaptive disclosure, and structured evidence notes; it reports strong LongMemEval performance while accessing only 8% of conversations.

1 Introduction

  • Relevant information can be scattered across long-term, multi-session interactions, leaving agents at risk of forgetting facts, evolving goals, or cross-session connections.The paper identifies this as contextual amnesia.Source: §1 · p. 1
  • APDMem stores conversations in four Markdown layers, from thematic summaries and key facts to turn-level evidence notes and raw messages.Raw messages remain beneath progressively abstract layers rather than being replaced by a specialized database, graph, or compressed-only representation.Source: §1 · pp. 1–2
  • Its controller treats memory access as query-adaptive disclosure, opening coarse views first and revealing finer evidence only when needed.It decomposes queries, handles temporal constraints, and drills deeper into the hierarchy when needed.Source: §1 · p. 2
  • The Organized Notes Synthesizer structures retrieved evidence and the controller’s draft answer into grounded notes for answer generation.These notes consolidate relevant evidence and align it with the query.Source: §1 · p. 2
  • APDMem achieves strong LongMemEval performance while accessing only 8% of conversations.The paper presents this as its LongMemEval validation result.Source: §1 · p. 2

2 Related Work

Prior work organizes conversational memory in several forms and uses finer-grained retrieval or evidence organization for complex queries. APDMem makes progressive disclosure an explicit, agent-controlled principle and structures retrieved evidence into notes for answer generation.

2.1 Memory Representations for LLM Agents

  • Memory systems span distilled profiles and facts, episodic traces, temporal graphs, and compressed semantic memories; APDMem instead uses transparent Markdown layers.The paper notes that existing approaches can require graph conversion, compression, or specialized memory objects.Source: §2.1 · p. 2

2.2 Progressive Disclosure

  • APDMem makes progressive disclosure an explicit architectural principle, with an agent controlling how deeply memory is accessed through tool-mediated interactions.It exposes coarse views first and finer evidence only when needed.Source: §2.2 · p. 2

2.3 Evidence-Grounded Generation

  • For complex queries, APDMem organizes retrieved evidence into notes that consolidate facts, temporal order, contradictions, enumeration evidence, and answer hints.The notes provide structured grounding context for the answer generator.Source: §2.3 · p. 2

3 APDMem

APDMem combines layered conversation memory, query-adaptive disclosure, and evidence synthesis to retrieve the detail needed for an answer. Its controller moves through the hierarchy as needed, while convergence rules bound retrieval and organized notes support final generation.

3 APDMem

  • APDMem combines a four-layer memory substrate, a progressive disclosure controller, and an organized-notes synthesizer for final answer generation.Source: §3 · p. 3

3.1 Multi-Resolution Memory Substrate

  • Each session stores thematic summaries, personalized key facts, turn-level evidence notes, and verbatim messages, with raw messages retained for exact evidence recovery.Source: §3.1 · p. 3

3.2 Progressive Disclosure Controller

  • The controller gathers sufficient query evidence while avoiding unnecessary detail, selecting both relevant memory regions and resolution levels with increasing cost and fidelity.Source: §3.2 · pp. 3–4

3.2.1 Controller Workflow

  • The workflow normalizes relative dates and decomposes compound queries, then searches summaries and drills through facts, ordered turns, and relevant raw messages as needed.Source: §3.2.1 · p. 4

3.2.2 Convergence Control

  • Retrieval stops when reflection finds no new evidence or the step budget is reached, bounding further controller iterations and tool calls.Source: §3.2.2 · p. 4

3.3 Organized Notes Synthesizer

  • The synthesizer organizes retrieved evidence and the controller’s draft into query-focused notes covering facts, temporal order, contradictions, counts, and an answer hint; generation uses these notes with the draft.Source: §3.3 · p. 4

4 Experiments

Experiments evaluate APDMem on 500 LongMemEval-S questions against Full-context, Mem0, LightMem, and SimpleMem using GPT-4.1 and GPT-4.1-mini backbones. The analyses examine accuracy, inference efficiency, disclosure behavior, ablations, and representative traces.

4.1 Main Experiments

  • 87.8% with GPT-4.1 and 79.8% with GPT-4.1-mini; APDMem exceeds SimpleMem by 3.8 and 2.9 points, respectively.The detail-oriented SS-Asst category reaches 100.0% with GPT-4.1, versus 76.8% for SimpleMem and 96.4% for Full-context.Source: §4.1 · pp. 4–5, Table 1 · p. 5

4.2 Inference Efficiency

  • APDMem uses roughly one-tenth the offline token budget of update-heavy systems and a small fraction of their latency, while avoiding aggressive compression with minimal latency.Baseline online costs are unavailable, so the comparison reports APDMem’s online controller cost separately.Source: §4.2 · p. 5, Table 2 · p. 5

4.3 Effectiveness of Progressive Disclosure

  • 8% of haystack sessions are inspected on average (3.7 of 47.7), with single-session questions requiring less access than multi-session questions.The reported rates are 2% for SS-Asst., 6% for SS-User, and 12% for multi-session questions; APDMem achieves 87.8% overall accuracy.Source: §4.3 · p. 5, Figure 2 · p. 6
  • 73.7% of 1,838 drills stop at L1, while 20.2% reach L3 and 6.1% stop at L2; L3 use is more frequent for single-session queries.L3 rates are 40.7% for SS-Asst., 35.8% for SS-User, 15.2% for multi-session, and 9.9% for knowledge-update questions.Source: §4.3 · p. 6, Figure 2 · p. 6
  • With a GPT-4.1-mini generator, upgrading the controller raises accuracy from 79.8% to 86.8%, while a GPT-5.2 controller reaches 87.8%.The error analysis reports retrieval errors falling from 22.2% to 10.6% as the backbone scales, while generation errors remain near 8–10%.Source: §4.3 · p. 6, Table 3 · p. 6, Figure 2 · p. 6

4.4 Ablation Studies

  • Removing the controller while retaining the hierarchy reduces accuracy from 87.8% to 83.2%; raw-message-only RAG scores 77.0%, a 12.3% relative drop.These comparisons distinguish the contributions of agent-controlled retrieval and the memory hierarchy.Source: §4.4 · pp. 6–7, Table 4 · p. 6

4.5 Case Studies

  • The traces show broadening across sessions at L1 for a multi-hop question, but descending from L1 to L3 within a session for a verbatim detail.In the first case, reflection prompts another search and the controller converges without descending; in the second, L3 recovers the shop name.Source: Figure 3 · p. 7, §4.5 · p. 7

4.4 Ablation Studies

  • Removing L1 facts or L3 raw messages causes the largest memory-level drops (3.6% and 5.9%), while removing L2 notes drops accuracy by 0.5%.Among reasoning tools, reflection contributes a 3.6% drop, decomposition 2.1%, and date interpretation 0.2%.Source: Table 4 · p. 6, §4.4 · p. 7

5 Conclusion

APDMem frames long-term memory access as query-adaptive disclosure over a lightweight hierarchy, with a controller exposing finer evidence as needed and a notes synthesizer organizing it for answer generation. On LongMemEval, it reaches 87.8% accuracy with GPT-4.1 and balances answer accuracy with token cost and latency.

5 Conclusion

  • APDMem combines a lightweight hierarchical memory, a controller that drills into finer evidence when needed, and a notes synthesizer for grounded answer generation.The conclusion identifies adaptive disclosure as the architecture’s organizing principle.Source: §5 · p. 7
  • On LongMemEval, APDMem reaches 87.8% accuracy with GPT-4.1; ablations identify agent-controlled progressive disclosure as a significant contributor to performance.Disclosure depth varies by question type: most questions use the key-fact layer, while some detail-oriented categories benefit from raw-message access.Source: §5 · p. 7
  • APDMem balances token cost and latency with strong answer accuracy, while using adaptive disclosure rather than a fixed retrieval granularity.The conclusion describes this as a practical efficiency–accuracy trade-off.Source: §5 · p. 7

Limitations

APDMem’s retrieval quality has boundaries tied to summary fidelity, first-access latency, and cross-session modeling. The paper describes mitigations or trade-offs for some of these constraints, while leaving cross-session linking for future work.

Limitations

  • L0 summaries can omit relevant details, lowering the corresponding turns’ BM25 and dense-retrieval scores.The keyword-field boost and reflect tool offer partial recovery, but initial ranking remains sensitive to summary fidelity.Source: Limitations · pp. 7–8
  • Lazy materialization of L1 and L2 may add cold-start latency on a session’s first access.Pre-materialization can reduce first-access latency, but may incur unnecessary cost for sessions or layers that are never queried.Source: Limitations · p. 8
  • APDMem does not explicitly model cross-session relationships as graph-based memory structures can.The paper identifies lightweight cross-session linking as future work.Source: Limitations · p. 8

Disclaimer

The appendix includes a disclaimer and documents the controller algorithm, system prompts, and experiment hyperparameters. The prompts span memory construction, retrieval and reasoning, and evidence organization and answer generation.

Disclaimer

  • The document disclaims investment research or advice, recommendations, and liability for the completeness, accuracy, or reliability of its information.It also states that the document is not a product of JPMorgan’s Research Department.Source: Disclaimer · p. 8

A Progressive Disclosure Controller Algorithm

B System Prompts

  • Figures 4–6 provide L0–L2 memory-construction prompts, while Figures 7–10 cover controller, reflection, notes synthesis, and answer generation.The construction prompts resolve pronouns to named entities and convert relative dates to absolute dates using the session timestamp; the synthesizer organizes evidence without answering, and the answer-generator prompt is shared with SimpleMem.Source: B System Prompts · pp. 9–10

C Hyperparameter Configuration

  • Memory-construction parameters are fixed across backbones, and retrieval parameters were selected on 20 held-out development questions and kept constant.The passage identifies L0 window size and L1 bullet bounds as the fixed memory-construction parameters.Source: C Hyperparameter Configuration · p. 10

D Setting the Controller Step Budget

Tmax caps the controller’s inference-time iteration and tool/LLM-call budget, bounding worst-case cost and latency. In experiments, nearly all queries stopped before the cap; application settings can trade additional search against resource use.

D Setting the Controller Step Budget

  • Tmax caps controller iterations and the additional tool/LLM calls before escalation must stop and the best available answer is returned.This makes it an explicit inference-time budget parameter.Source: D Setting the Controller Step Budget · p. 11
  • With Tmax = 20, nearly all experimental queries terminated before reaching the cap, which still bounded worst-case compute.The passage says the budget did not materially affect answer quality for most examples.Source: D Setting the Controller Step Budget · p. 11
  • Larger Tmax permits continued search and reflection on difficult queries, while smaller values reduce worst-case latency and tool/LLM usage.The paper recommends profiling iteration demand, respecting operational bounds, and choosing the tightest sufficient cap.Source: D Setting the Controller Step Budget · p. 11

E Statistical Significance Analysis

Across 10 runs on LongMemEval-S with GPT-4.1, APDMem outperformed SimpleMem on mean accuracy, and the reported confidence intervals place the gap above zero. The appendix also reports high human–judge agreement and describes the hybrid retrieval procedure.

E Statistical Significance Analysis

F Human Validation of the LLM-as-a-Judge

G Hybrid Retriever

  • The Hybrid Retriever searches L0 thematic-summary units using complementary lexical and dense indexes, then combines their turn rankings with Reciprocal Rank Fusion.The lexical index uses BM25; the dense index embeds summaries, and the two rankings receive equal RRF weights.Source: G Hybrid Retriever · p. 13
  • For decomposed questions, sessions are scored by their best-matching turn across queries, with an exact-match keyword-field boost to preserve lexical precision.The top-K sessions are returned with their thematic summaries.Source: G Hybrid Retriever · p. 13

E Statistical Significance Analysis

  • Both confidence intervals for the APDMem–SimpleMem accuracy gap are above zero, indicating a statistically significant improvement with GPT-4.1.The analysis uses Welch’s t-test and a bias-corrected, accelerated bootstrap interval.Source: E Statistical Significance Analysis · p. 12, Table 8 · p. 14

H Per-Model Efficiency Breakdown

Per-model efficiency varies across offline construction and online retrieval, while the memory substrate stores conversation history in four layers of increasing detail. The layers range from precomputed thematic summaries to verbatim messages, with intermediate facts and turn-level evidence generated on demand.

H Per-Model Efficiency Breakdown

  • Offline construction totals 145–154k tokens across models, with output tokens rising from 15.2k for GPT-4o-mini to 23.7k for GPT-5.2.All models share the same 130.1k-token input cost because L0 processes identical conversation windows.Source: H Per-Model Efficiency Breakdown · p. 14
  • GPT-5.2 uses 86.5k tokens and has 42.7s latency, while GPT-4.1-mini uses 86–89k tokens but has the highest latency at 57.7s.The passage attributes the stronger models’ lower online token use to fewer ReAct steps, better initial search queries, and earlier identification of evidence gaps.Source: H Per-Model Efficiency Breakdown · p. 14

I Multi-Resolution Memory Substrate: Construction and Examples

  • The substrate construction section presents examples from the same user’s 44 sessions and identifies prompts for L0, L1, and L2.The prompts are shown in Figures 4, 5, and 6, respectively.Source: I Multi-Resolution Memory Substrate:… · p. 14

I.1 L0: Thematic Summaries

  • L0 is materialized at ingestion: overlapping four-turn windows produce one-sentence summaries, keywords, and tags, with pronouns resolved and relative dates anchored to session timestamps.The summaries are embedded for dense retrieval; Figure 11 shows an example output.Source: §I.1 · p. 14, Figure 11 · p. 15

I.2 L1: Personalized Key Facts

  • L1 is generated on demand when read_facts is called, producing 5–25 self-contained facts about user information, preferences, plans, and constraints, then caching them.Figure 12 shows an example output.Source: §I.2 · p. 14, Figure 12 · p. 15

I.3 L2: Turn-Level Evidence Notes

  • L2 is generated on demand by splitting turns into separate, self-contained fact bullets that preserve specifics such as names, dates, and quantities, then caching them.Figure 13 shows an example of atomic fact bullets grouped by turn.Source: §I.3 · p. 15, Figure 13 · p. 15

I.4 L3: Raw Messages

  • L3 stores verbatim user–assistant messages by turn without LLM processing; read_raw accesses them when higher layers lack exact wording or fine-grained details.Figure 14 shows an example output.Source: §I.4 · p. 15

J Generator-Side Ablation

With retrieval held fixed, the generator-side ablation tests Organized Notes and the controller’s draft answer. The full system scores 87.8%, and removing either component—or both—reduces accuracy.

J Generator-Side Ablation

  • 87.8% is the full system’s accuracy; removing the draft answer lowers it by 1.2 points, removing Organized Notes by 2.6 points, and removing both yields 83.6% (−4.2).The ablation fixes retrieval results from one run and varies only the generation strategy.Source: J Generator-Side Ablation · p. 15
  • Organized Notes improves multi-session synthesis by 6.0 points over raw evidence and temporal reasoning by 3.0 points, while the draft benefits SS-User by 2.9 and SS-Assistant by 1.8.Organized Notes organizes facts by relevance, enumerates items, and highlights contradictions; the draft supplies focused controller reasoning for single-session categories.Source: J Generator-Side Ablation · p. 16

Related papers

Loading 2610.02472v1…