PaperlayerView exact revision on arXiv

Source-linked AI summary

CABLE: Extending the Reach of Memory Retrieval via Complementary Antecedent-Based Linking and Expansion

Zheling Tan, Jin Gao, Dequan Wang

arXiv:2608.17911v1cs.CL

TL;DR

Long-term memory retrievers can miss earlier experiences, plans, or motivations that explain later events when those memories fall outside the direct semantic neighborhood. CABLE adds sparse, verified antecedent links that complement semantic retrieval, yielding higher overall mean LLM-judge scores across every evaluated system-level setting, especially when evidence spans memories or sessions.

  • Problem

    Bounded semantic retrieval can miss earlier experiences, plans, motivations, or background events needed to answer later queries.

  • Method

    CABLE constructs sparse verified antecedent links after removing direct semantic overlap, then expands retrieved seeds along non-redundant graph edges.

  • Results

    CABLE yields higher overall mean LLM-judge scores in every evaluated system-level setting, with strongest gains when evidence is distributed across memories or sessions.

  • Takeaways & Limitations

    Retriever-complementary associations provide a practical mechanism for improving evidence reachability in long-term conversational memory.

  • Takeaways & Limitations

    CABLE’s append-only graph does not remove or revise verified edges after insertion, so the graph can continue growing over an agent’s lifetime.

Abstract

from arXiv · show

As LLM agents operate across structured workflows and sessions, preserving long-term history does not ensure that later contexts can recover relevant evidence through a bounded memory interface. We study this evidence-reachability problem in long-term conversational memory, where retrieval still relies heavily on semantic similarity. This works well for topical recall, but it often misses earlier experiences, plans, or motivations that are semantically distant from the later events they help explain. Existing memory graphs provide cross-memory structure, yet links driven mainly by semantic overlap can duplicate what the host retriever already recovers. We argue that link construction should instead prioritize a sparse set of retriever-complementary associations. We present CABLE (Complementary Antecedent-Based Linking and Expansion), a plug-in augmentation that constructs links designed to extend the host retriever's direct semantic reach. For each new memory, CABLE generates antecedent-oriented queries, retrieves prior memories, subtracts candidates in the direct semantic neighborhood, and verifies the remainder before adding the accepted complementary associations into a sparse directed graph. At retrieval time, CABLE expands the host system's retrieved seeds along these links to surface implicit supporting evidence. We evaluate CABLE with A-MEM on LoCoMo and MA-LongMemEval, and further integrate it into SimpleMem and Mem0g on LoCoMo, using Qwen3.5-27B, DeepSeek-chat, and GPT-4o-mini. CABLE yields higher mean LLM-judge scores in every evaluated system-level setting, with the largest gains in categories where useful evidence is distributed across memories or sessions, including open-domain, multi-session, and preference-oriented questions. These results support prioritizing sparse, reasoning-relevant associations that complement rather than duplicate the host retriever.

1 Introduction

Long-term memory systems can store relevant evidence that falls outside a later query’s direct semantic neighborhood, limiting recovery through bounded retrieval. CABLE addresses this reachability problem with sparse, retriever-complementary antecedent links and improves overall mean LLM-judge scores across evaluated system-level settings.

  • Problem: Execution graphs define computation and state transitions but do not specify how later components recover relevant history across component and session boundaries.This creates an information-access problem when history no longer fits inside one prompt.
  • Problem: Semantic retrieval often misses earlier experiences, plans, motivations, or background events that explain a later query while remaining outside its direct semantic neighborhood.The required evidence may be stored but inaccessible within the retriever’s limited output budget.
  • Design principle: Retriever-complementary associations should expose useful evidence beyond direct retrieval rather than duplicate semantic overlap, while remaining sparse, persistent, and inspectable.This shifts structured memory toward providing marginal retrieval access instead of representing relations for their own sake.
  • Method: CABLE constructs antecedent-oriented links during memory writing by retrieving candidate prior memories, subtracting the direct semantic neighborhood, and verifying the remaining associations.The links are reused during retrieval-time expansion, which requires no additional LLM calls.
  • Evaluation: CABLE produces higher overall mean LLM-judge scores in every evaluated system-level setting across A-MEM, SimpleMem, and graph-enabled Mem0g experiments.Evaluations use LoCoMo and MA-LongMemEval with Qwen3.5-27B, DeepSeek-chat, and GPT-4o-mini; gains are largest for open-domain, multi-session, and preference-oriented questions.

2 Related work

Related work develops long-term memory through explicit entry-based storage, relational memory structures, and retrieval methods that construct intermediate query-side representations. These approaches strengthen memory maintenance, associative navigation, and evidence retrieval beyond direct item-level matching.

  • Entry-based memory systems: Entry-based memory systems store past interactions as discrete memory entries, differing mainly in how those entries are formed and maintained.MemoryBank models long-term memory as evolving user-related records with forgetting-aware updates for temporal decay.
  • Structured memory association: Structured memory association adds relational organization to improve associative access beyond largely independent memory entries.A-MEM uses dynamic indexing and linking, while CompassMem segments experience into events connected through explicit logical relations in an Event Graph.
  • Structured memory association: Some structured-memory methods combine enriched links with reasoning-time control through causal, semantic, multi-view, or policy-guided graph operations.ActMem integrates retrieval with counterfactual reasoning, MAGMA uses multiple relational views and policy-guided traversal, and Hindsight supports temporal and entity-based structure.
  • Retrieval beyond direct matching: Retrieval beyond direct matching constructs intermediate query-side representations to improve how stored evidence is retrieved.HyDE generates a hypothetical document from the query before retrieving evidence using that representation.

3 Method

CABLE constructs sparse directed links from prior memories to each new memory by generating antecedent-oriented queries, removing semantic overlap, and verifying candidates. At retrieval time, it expands the host retriever’s seeds along these links to expose alternative access paths while filtering redundant additions.

  • Candidate selection: CABLE retrieves direct semantic neighbors and antecedent candidates, subtracts their overlap, and verifies the remaining candidates before storing links.This procedure is designed to retain non-redundant associations that complement the host retriever.
  • Retrieval-time expansion: At query time, CABLE expands the host system’s retrieved seeds along stored links and filters redundant additions.The expansion is intended to surface implicit supporting evidence through links constructed during memory formation.
  • Link construction: CABLE creates directed edges from prior memories to each arriving memory, providing alternative access paths that semantic retrieval is unlikely to surface directly.The links are intentionally sparse and target complementary evidence rather than direct semantic neighbors.
  • Antecedent query generation: For each new memory, CABLE generates antecedent-oriented queries about earlier experiences, plans, motivations, or background events.A rule-based filter removes low-information memories, after which an LLM classifies each memory operationally and generates a small query set guided by its type.

3. Subtract Overlap · 1. Antecedent-Oriented Query

CABLE constructs sparse, retriever-complementary antecedent links by querying for prior causes, motivations, and context, subtracting direct-search overlap, and verifying the remainder. At retrieval time, it expands reliable host-retrieval seeds through these links while filtering near-duplicates.

  • 1. Antecedent-Oriented Query: CABLE generates antecedent-oriented queries for prior experiences, causes, motivations, and contextual explanations of each new memory.For example, “Emma submitted her resignation letter” prompts queries about work pressure, colleague conflict, and salary and benefits.
  • 3.1.2 Dual retrieval: CABLE performs parallel direct and antecedent searches, collecting top-Kb direct neighbors and the union of top-Kh results across antecedent queries.Repeated antecedent-search results retain their highest score.
  • 3.1.3 Overlap subtraction: CABLE defines the complementary candidate set as Ci = Hi \ Bi, excluding any memory already returned by direct semantic retrieval.If Ci = ∅, CABLE adds no edge for mi.
  • 3.1.4 Verification and graph update: An LLM verifier accepts candidates supplying prior causes, motivations, enabling events, background, or earlier states, while rejecting topical or entity overlap alone.This prevents weak or coincidental associations from unnecessarily densifying the graph.
  • 3.1.4 Verification and graph update: Each accepted antecedent association adds a directed graph edge mj → mi from the prior memory mj to the new memory mi.The graph stores only verified links.
  • 3.2 Retrieval-stage extension: At query time, CABLE first preserves the host system’s initial result set R0, then expands it through graph-based seed selection, candidate scoring, and novelty filtering.Only seeds meeting the cosine-similarity threshold τ are retained.
  • 3.2.1 Seed selection and candidate scoring: One-hop incoming links surface earlier memories explaining retrieved events, while outgoing links surface later events explained by retrieved causes or plans.Candidate scores aggregate support from connected seeds, favoring candidates supported by multiple high-confidence seeds.
  • 3.2.2 Novelty filtering: CABLE greedily selects candidates within the expansion budget and accepts them only when sufficiently novel relative to the current result set R.This preserves the host retriever’s backbone while adding non-redundant graph expansions.

4 Experiments

CABLE is evaluated on LoCoMo and MA-LongMemEval across A-MEM, SimpleMem, and Mem0g using mean LLM-judge scores. It improves system-level performance broadly, especially when evidence spans sessions or encodes preferences, while overlap subtraction and verification support its effectiveness.

  • Benchmarks: LoCoMo contains 1,540 answerable questions across Single-hop, Multi-hop, Temporal Reasoning, and Open-domain Knowledge categories.The categories emphasize direct retrieval, multi-memory composition, temporal dependencies, and broader user- or world-aware reasoning.
  • Systems and setup: CABLE is integrated into A-MEM, SimpleMem, and Mem0g, with controlled comparisons that replace retrieved entries rather than increasing the total context.A-MEM and Mem0g replace up to five entries; A-MEM passes at most 45 entries, while Mem0g retrieves 20 memories in total.
  • A-MEM results: +6.00 percentage points is CABLE’s largest A-MEM system-level gain, observed on MA-LongMemEval with Qwen3.5-27B.The other reported A-MEM gains are +3.58 on LoCoMo with Qwen3.5-27B, +2.11 on LoCoMo with DeepSeek-chat, and +1.00 with GPT-4o-mini.
  • Question-type results: +23.33 is CABLE’s largest reported category gain, improving single-session-preference on MA-LongMemEval with Qwen3.5-27B.Multi-session also improves by +12.00 under Qwen3.5-27B; on LoCoMo, Open-domain gains reach +6.24 with Qwen3.5-27B and +9.37 with DeepSeek-chat.
  • Cross-system results: CABLE improves SimpleMem by +0.58 and +1.62, and improves Mem0g from 52.6% to 54.8% while keeping the retrieved-entry count fixed.Mem0g gains appear across Single-hop (+2.0), Multi-hop (+2.5), Temporal (+1.9), and Open-domain (+4.2) questions.
  • Ablations: Overlap subtraction and LLM verification both contribute to CABLE, while type-conditioned antecedent queries outperform generic query decomposition by 0.46 percentage points on LoCoMo.Removing overlap subtraction allows redundant links to consume the expansion budget, whereas removing verification retains noisier associations.

5 Conclusion · A Integration protocols

CABLE addresses evidence-reachability by adding verified antecedent links that complement direct semantic retrieval. Across benchmarks, memory systems, and LLMs, it produces consistent gains while preserving each host system’s native retrieval regime.

  • 5 Conclusion: CABLE addresses bounded-interface evidence recovery at the association-and-retrieval layer.It constructs and reuses verified antecedent links to extend access to relevant historical evidence.
  • 5 Conclusion: CABLE’s antecedent links are designed to complement rather than duplicate direct semantic retrieval.The links target associations that direct semantic retrieval may not recover.
  • 5 Conclusion: Experiments across two benchmarks, three memory systems, and multiple LLMs show consistent overall gains.The conclusion reports this pattern across the evaluated settings without specifying a single aggregate value.
  • 5 Conclusion: The strongest improvements occur when useful evidence is distributed across memories or sessions.These conditions motivate linking evidence that is not directly adjacent to the later query.
  • A Integration protocols: CABLE integration preserves each host system’s native retrieval regime.The protocols are adapted to the retrieval behavior of A-MEM, Mem0g, and SimpleMem.
  • A Integration protocols: A-MEM and Mem0g use matched fixed retrieved-entry budgets, while SimpleMem retains adaptive retrieval.This protocol alignment supports comparisons without replacing each host system’s retrieval strategy.
  • A Integration protocols: SimpleMem invokes CABLE only after insufficient baseline retrieval.CABLE is therefore applied conditionally within SimpleMem’s adaptive reflection-based protocol.

B Construction Cost, Graph Growth, and Retrieval-Time Overhead

CABLE shifts extra computation to incremental link construction, where costs depend on overlap-surviving candidates rather than the full memory base. At retrieval time, it adds no LLM calls and expands selected seeds through graph neighborhoods using embeddings and novelty filtering.

  • Construction cost: CABLE performs one classification call, one query-generation call, and one verification call per overlap-surviving candidate for each new memory.Construction is incremental and does not recompute the graph over previously stored memories.
  • Graph growth: Overlap subtraction and verification prevent redundant or weak links, allowing CABLE to avoid all-pairs verification.Verification rejects topical co-occurrence, entity overlap, and non-substantive memories.
  • Graph growth: Each new memory generates at most Nq antecedent-oriented queries and each query retrieves at most Kh candidates, limiting candidate consideration to at most NqKh antecedents.The passage also states that stored links are bounded by the same product, but the supplied text truncates the full expression.
  • Retrieval-time overhead: CABLE adds no LLM calls at retrieval time, instead enumerating one-hop neighborhoods, scoring candidates with available embedding similarities, and applying novelty filtering.These operations are performed for the selected seed set S.
  • Graph growth: The global bound |E| = O(|M|) does not guarantee constant degree for every seed, while integrations differ in how retrieved-entry counts and adaptive protocols are controlled.A-MEM and Mem0g keep retrieved-entry counts fixed; SimpleMem retains its host-controlled adaptive protocol.

C Component and Query-Design Ablations

Component ablations on A-MEM show that overlap subtraction and verification improve performance across benchmarks, while type-conditioned antecedent queries outperform generic topical decomposition on LoCoMo.

  • Experimental setup: Table 5 evaluates CABLE component ablations on LoCoMo and MA-LongMemEval, while generic query-decomposition comparisons are limited to LoCoMo.The experiments use A-MEM with Qwen3.5-27B.
  • Component ablations: 0.91 percentage points on LoCoMo and 0.67 points on MA-LongMemEval are lost when overlap subtraction is removed.This supports retriever-complementary rather than redundant links.
  • Component ablations: Removing verification causes a smaller reduction on LoCoMo but a 2.66-point reduction on MA-LongMemEval.Longer histories can introduce noisier antecedent candidates, increasing the value of verification.
  • Query design: Replacing type-conditioned antecedent queries with generic topical decomposition reduces the LoCoMo score.The supplied passage reports the direction of this effect but not its magnitude.

D Qualitative case study · E Prompts

The qualitative case study shows CABLE recovering motivation-related evidence that semantic retrieval misses, while the prompt pipeline operationalizes memory typing, antecedent-query generation, link verification, and evaluation.

  • D Qualitative case study: CABLE retrieves an absent memory explaining why Melanie used colors and patterns, whereas the baseline only establishes her enjoyment of pottery.The expanded memory states that Melanie wanted the colors and patterns to catch the eye and make people smile.
  • E Prompts: The prompts cover four pipeline stages: memory typing, antecedent-query generation, link verification, and evaluation.Runtime slots include memory content, participant lists, and expected output schemas.
  • E.1 Memory type classification: Each memory is assigned one of four operational types, with temperature 0.1 decoding and keyword-classifier fallback when parsing fails.The types are opinion, event, plan, and state_change.
  • E.2 Antecedent-oriented query generation: CABLE generates at most Nq = 3 antecedent-oriented queries from the assigned type, using temperature 0.2 decoding and fallback templates when classification fails.If neither the LLM nor rule-based classifier yields a supported type, the fallback query template is used.
  • E.3 Link verification: Every candidate is verified at temperature 0.0 for useful background or implicit evidence before an edge is inserted.Verification targets non-obvious links such as distant causes, explanatory context, character development, implicit motivation, and multi-hop evidence.
  • E.3 Link verification: The verifier rejects non-substantive memories, unrelated coincidences, and connections lacking a plausible reasoning chain.It returns a JSON validity decision with a one-sentence reason.
  • E Prompts: The LLM judge scores generated answers against the question and reference answer, retaining intermediate scores without rounding or binarization.Reported results are mean LLM-judge scores expressed as percentages, with baseline and +CABLE systems using the same prompt and judge.

F Limitations · G Use of LLMs

CABLE is limited by append-only graph growth, retrieval-expansion costs, and evaluation confined to conversational-memory question answering. The authors also used LLMs throughout development and evaluation, with all generated suggestions critically reviewed and revised.

  • F Limitations: CABLE augments retrieval without redesigning memory writing, updating, consolidation, or discarding, which remain the host system’s responsibility.Its current link construction is append-only, making each verified association a persistent edge without later removal or revision.
  • F Limitations: Append-only associations can cause the graph to grow continually as verified edges accumulate.CABLE currently has no mechanism for removing or revising inserted edges.
  • F Limitations: Retrieval-time expansion visits selected seeds’ one-hop neighborhoods, so its cost may increase with accumulated association degrees.The computational burden depends on the degrees of the selected seeds.
  • F Limitations: Persistent edges can surface evidence after later memories supersede or correct it, motivating forgetting, decay, or consolidation policies.Integrating such policies with CABLE is identified as an important future direction.
  • F Limitations: Experiments evaluate CABLE only on long-term conversational-memory question answering, excluding tool-using agents, multi-agent coordination, and graph-level task execution.Testing CABLE in these broader workflows remains future work.
  • G Use of LLMs: LLMs supported memory extraction, link construction, answer generation, LLM-as-judge evaluation, and manuscript framing, restructuring, and language revision.The authors critically reviewed, verified, and revised all LLM-generated suggestions and accepted responsibility for the reported results and manuscript content.
Loading 2608.17911v1…