PaperlayerView exact revision on arXiv

Source-linked AI summary

ActionRAG: Evidence-Refining Graph Diffusion for Long-Term Memory in LLM Agents

Navyansh Mahla, Sashank Vandrangi

arXiv:2610.32752v1cs.AIcs.CL

TL;DR

Flat retrieval struggles with multi-hop relations and superseded facts, motivating memory that can reconcile evidence across an interaction history. ActionRAG builds a cautiously indexed graph and refines query-local evidence; across MAB’s deterministic columns, it outperforms retrieval-based baselines on most columns. Its reported timing comparison is hardware-mismatched.

  • Problem

    Flat chunk retrieval scores facts independently, making multi-hop relations and superseded facts difficult to handle in memory for agents.

    Source: §1 · pp. 1–2

  • Method

    ActionRAG builds a cautiously indexed entity-relation graph, retrieves entity and triple seeds, and uses query-conditioned diffusion to refine evidence in a bounded local subgraph.

    Source: §1 · p. 2, §3.1 · pp. 4–5, §3.2 · p. 5

  • Results

    Across eight heterogeneous deterministic MAB columns, ActionRAG has the highest reported retrieval-based score on six and an unweighted mean of 55.1, 5.4 points above HippoRAG-v2.

    Source: §1 · p. 2, §4.2 · p. 9

  • Takeaways & Limitations

    The best reported configuration combines reconciled, summarized indexing with retrieval-time refinement.

    Source: §4.2 · p. 9

    Timing results are not hardware-matched: ActionRAG ran on consumer Apple Silicon, while reported HippoRAG-v2 timings used four NVIDIA L40 GPUs and an AMD EPYC 7713 CPU.

    Source: ActionRAG configuration · p. 9

Abstract

from arXiv · show

LLM agents need long-term memory that can recover precise and mutually consistent evidence from growing interaction histories, particularly when answers depend on relations among facts rather than isolated semantic matches. Current memory pipelines remain vulnerable at two stages: indexing errors can fragment repeated entities or merge distinct ones, and online retrieval often treats initial similarity scores as final even when the surrounding graph provides evidence for revising them. We introduce ActionRAG, a graph memory architecture that couples cautious indexing with query-time graph retroevaluation. Its indexing pipeline reconciles type labels and applies tiered entity resolution. Given a query, dual-channel entity and triple retrieval forms an initial evidence set and induces a bounded local graph. Query-conditioned receiver-normalized diffusion refines triple scores using local path structure. Seed protection, hub-penalized flux, and triple materialization preserve direct evidence while recovering missed connections. We evaluate ActionRAG on MemoryAgentBench (MAB), which tests accurate retrieval, conflict resolution, long-range understanding, and test-time learning. Across its eight deterministic-scoring columns, ActionRAG achieves 55.1 average accuracy, improving over the strongest retrieval-based memory baseline by 5.4 points. Under MAB's query-time cost accounting, mean synthesis cost stays below \$0.001 per query on the four reported cost columns, and the sum of separately measured retrieval-side p99 components is 440.7 ms on our consumer Apple Silicon setup.

1. Introduction

MemoryAgentBench evaluates memory across accurate retrieval, test-time learning, long-range understanding, and conflict resolution, while flat chunk retrieval struggles with multi-hop relations and superseded facts. ActionRAG combines cautious graph indexing with query-local evidence refinement and leads retrieval-based systems on six of eight deterministic MAB columns.

1. Introduction

  • MemoryAgentBench assesses accurate retrieval, test-time learning, long-range understanding, and conflict resolution across incremental multi-turn interactions.These competencies require recalling specific facts, reasoning across facts, recognizing superseded statements, and absorbing new information.Source: §1 · p. 1
  • Flat chunk retrieval scores each chunk independently, weakening multi-hop reasoning and making superseded facts hard to distinguish from the facts that override them.The two statements may share most of their semantic content.Source: §1 · pp. 1–2
  • Graph-based alternatives use shallow traversal or entity-seeded Personalized PageRank, while ActionRAG retrieves entity and triple seeds directly to preserve relational cues without query-time LLM parsing.Entity-only seeding may overlook predicate-level cues that are not salient named entities.Source: §1 · p. 2
  • 55.1 unweighted mean across MAB’s eight deterministic columns, 5.4 points above HippoRAG-v2, with the highest reported retrieval-based score on six columns.The result combines reconciled, summarized indexing with query-local refinement.Source: §1 · p. 2
  • ActionRAG's contributions include cautious indexing with label reconciliation, tiered entity resolution, provenance preservation, and conflict-aware summarization.Its local graph diffusion combines dual-channel retrieval with relational retroevaluation.Source: §1 · p. 2

2. Related Work

Related work ranges from extending context windows and external memory systems to retrieval pipelines and graph-based memory. These approaches trade off context length, structural signal, query-time computation, and reliance on selected entity cues.

2.1. Long-Context and Memory-Augmented LLMs

  • Long-context approaches extend windows but lose accuracy for deep information and raise inference cost; external memories store, structure, or actively manage interaction history.External-memory examples include working-memory paging, learned retrieval, and memory graphs.Source: §2.1 · pp. 2–3

2.2. Retrieval-Augmented Generation

  • Dense retrieval is competitive for accurate retrieval on MemoryAgentBench but underperforms on multi-hop and conflict-resolution questions because independent chunk scoring misses cross-chunk structure.RAG systems may add re-rankers or late-interaction models.Source: §2.2 · p. 3

2.3. Knowledge Graphs for Retrieval and Reasoning

  • Graph methods span summaries, dynamic links, memory-guided clues, and global diffusion, trading off query-time calls, graph-wide computation, and dependence on entity cues; Figure 1 depicts ActionRAG's local retrieval pipeline.The figure shows entity anchors, concrete triple evidence, local diffusion, pruning, and materialization.Source: §2.3 · pp. 3–4, Figure 1 · p. 4

3. ActionRAG

ActionRAG separates graph construction from query-local retrieval, assuming relevant evidence lies in a bounded neighborhood. It combines cautious indexing with dual-channel retrieval and graph diffusion to refine evidence and recover missing connections.

3.2. Cautious Graph Construction

  • Chunk extraction emits typed entities and triples; types are reconciled across the batch before graph updates, and temporal phrases remain in descriptions.When a session timestamp is available, relative dates are converted to absolute dates; graph items retain source-chunk links.Source: §3.2 · pp. 5–6
  • After type canonicalization, compatible exact names are reused, fuzzy matches only when clearly best, and remaining mentions grouped before relations are written.Repeated observations merge into the same edge, with each graph item linked to its source chunks.Source: §3.2 · pp. 5–6
  • Repeatedly supported nodes and relations receive hierarchical summaries that retain dates and source links, keeping compact descriptions connected to original evidence.Source: §3.2 · p. 6

3.3. Query-Time Evidence Refinement by Graph Diffusion

  • ActionRAG treats initial retrieval as a proposal, building a bounded query-local graph and refining item scores through receiver-normalized diffusion.The bounded-neighborhood assumption keeps diffusion on a query-conditioned subgraph rather than the corpus-level graph.Source: §3.1 · pp. 4–5, §3.3 · pp. 6–7
  • Separate nearest-neighbor searches over entity and concrete-triple descriptions capture referential and predicate matches; bounded expansion forms an initial set of triples or orphan entities.Online relation retrieval means retrieval of concrete triples, not relation types.Source: §3.1 · p. 5, §3.3 · pp. 6, 8
  • Items interact through short, query-relevant paths: traversal cost selects a bounded path, hop length decays affinity, and salient triples can support diffusion across context-carrier edges.Source: §3.3 · pp. 6–7
  • Receiver-normalized diffusion combines each item's semantic evidence with a relevance-weighted neighbor average, while pruning protects strong raw-similarity items.The retained set includes items above a fraction of the maximum refined score and the strongest raw-score items.Source: §3.3 · p. 7
  • Hub-penalized flux ranks unseen triples between retained items for budgeted materialization; retained items are packed first, followed by bridge triples, under a token budget.Materialized evidence is temporary to the query, and retrieval invokes no LLM.Source: §3.3 · p. 7

4. Experiments

ActionRAG is evaluated on MemoryAgentBench’s eight deterministic columns, spanning four memory competencies. It leads retrieval-based systems on six columns, while component and cost analyses examine its indexing, retrieval, and query-time behavior.

4.1. Setup

  • The benchmark reports eight columns across accurate retrieval, conflict resolution, long-range understanding, and test-time learning, using MAB’s within-column aggregation.The deterministic average is the unweighted mean of the eight column scores; tasks requiring an LLM judge are excluded.Source: §4 · p. 8, §4.1 · pp. 8–9

ActionRAG configuration.

  • Each query retrieves ten entity and ten triple seeds, uses one receiver-normalized solve, and packs graph facts into a 7,000-token context.Indexing uses Claude Haiku 4.5 for extraction and graph summarization; ActionRAG ran on consumer Apple Silicon, unlike the hardware used for HippoRAG-v2 timings.Source: ActionRAG configuration · p. 9

4.2. Main results on MAB deterministic columns

  • 55.1 unweighted mean, 5.4 points above HippoRAG-v2, with the highest reported retrieval-based score on six MAB columns.ActionRAG trails HippoRAG-v2 on MH-DocQA and Detective QA. On MCC, it scores 82.0 versus 82.4 for GPT-4o-mini long context.Source: §4.2 · p. 9, Table 1 · p. 10, Table 3 · p. 11
  • 55.1 on the reconciled graph compares with 45.5 on the raw graph and 48.9 with summarization without reconciliation.Base retrieval on the reconciled graph scores 45.0; the best result combines reconciled, summarized indexing with query-local refinement.Source: §4.2 · p. 9, Table 2 · p. 10

4.3. Amortized per-query cost

  • ActionRAG costs at most $0.00116/query; separately measured retrieval and encoding components sum to 440.7 ms at p99.The cost accounting excludes graph construction, storage, and local embedding.Source: §4.3 · pp. 10–11, Table 4 · p. 11

5. Conclusion

ActionRAG combines reconciled graph construction with query-local receiver-normalized diffusion, pruning, and materialization. Across eight deterministic MAB columns, it leads retrieval-based systems on six, with component ablations favoring reconciled, summarized indexing plus retrieval-time refinement.

5. Conclusion

  • ActionRAG combines reconciled graph construction with query-local receiver-normalized diffusion, pruning, and materialization for long-term agent memory.The conclusion identifies more robust graph-expansion algorithms as future work.Source: §5 · p. 11
  • Across eight heterogeneous deterministic MAB columns, ActionRAG records the highest reported retrieval-based score on six and an unweighted mean of 55.1.The conclusion notes weaker MH-DocQA and Detective QA results alongside gains centered on test-time learning and conflict resolution.Source: §5 · p. 11
  • The conclusion reports low query cost and sub-half-second main-evaluation retrieval, while packed context remains nearly constant as the graph grows.Future work is directed toward graph-expansion algorithms that are more robust to hyperparameter values.Source: §5 · p. 11

Appendix A. Retrieval top-k ablation detail

Table 5 reports top-k accuracy across settings indexed by k.

Appendix A. Retrieval top-k ablation detail

  • Table 5 compares top-k accuracy, with k denoting the MAB baseline chunk count and KE=KT=k for the other setting.The supplied caption does not state the accuracy values or identify a winning setting.Source: Appendix A · p. 14, Table 5 · p. 15

Appendix B. Additional Ablations and Sensitivity Analyses

The appendix reports MCC sensitivity to retrieval settings and embeddings, alongside EventQA graph-growth measurements.

MCC sensitivity

  • MCC means span 82.0–84.8; α = 0.85 scores 84.8 and L = 0 scores 83.4, both above the fixed default.The fixed default is not MCC-optimal, materialization is not uniformly beneficial, and all settings remain within 2.8 points.Source: Appendix B · p. 15

MCC embedding sensitivity

  • On a fresh MCC run, BGE-M3 scores 84.0, above text-embedding-3-large at 81.0 and text-embedding-3-small at 79.6.Each backend uses a fresh run after rebuilding embeddings, separate from the main-table run; TREC-F varies most.Source: Appendix B · p. 16

EventQA graph growth

  • From 64K to full EventQA, graph nodes grow from 409 to 2,948 and tokens 8.4×, while packed context stays near 6.9K.Latency rises 25%/32% at p50/p99, and mean score falls 4.2 points as distractors increase.Source: Appendix B · p. 16, Table 8 · p. 16

Appendix C. ActionRAG configuration

The fixed ActionRAG configuration combines a specified extraction, embedding, and synthesis stack with one query-time diffusion solve.

Appendix C. ActionRAG configuration

  • All MAB columns use Claude Haiku 4.5 for extraction and summarization, BGE-M3 embeddings, and gpt-4o-mini synthesis.Table 9 presents the fixed retrieval configuration and maps implementation settings to Section 3.3.Source: Appendix C · p. 16, Table 9 · p. 16
  • The implementation uses cτ = max(0.01, 1 −cos(z, ϕ(τ))) and performs one diffusion solve.In per-intermediary mode, unseen path triples receive pairwise flux, duplicates are merged by maximum Fτ, and the top L become temporary query-local evidence.Source: Appendix C · p. 16

Appendix D. Retrieval latency protocol

Retrieval latency is measured separately from answer generation on a local Apple Silicon setup, with component timings reported independently.

Appendix D. Retrieval latency protocol

  • Isolated retrieval excludes answer generation and is measured on an Apple Silicon laptop using local Postgres/pgvector without a discrete GPU.Diffusion timing covers 50 MH-DocQA queries after initialization, including graph construction through context packing but excluding query embedding.Source: Appendix D · p. 16
  • BGE-M3 embedding latency is measured separately at batch size one, with 20 warm-ups and 200 measured calls.Embeddings are normalized.Source: Appendix D · p. 16
  • The reported p99 adds the retrieval and embedding component p99 values; concurrent full-retrieval wall time is omitted because database and Python scheduling are more variable.Source: Appendix D · p. 16
Loading 2610.32752v1…