RippleMem: From Isolated Retrieval to Associative Recollection for Long-Term Agent Memory

Abstract: LLM-based agents increasingly rely on external memory to support long-horizon reasoning and interaction. However, the main bottleneck is not simply storing past experience, but recovering the right set of evidence when relevant information is distributed across many interactions. Existing approaches struggle with this access problem. Full-context methods require noisy long-context search, flat retrieval often returns isolated and incomplete records, and graph-based memory systems can be expensive to construct while compressing rich event context. We introduce RippleMem, a long-term memory system that replaces one-shot retrieval with adaptive associative recollection. Inspired by cue-dependent episodic retrieval and associative completion, RippleMem stores interaction history as cue-rich episodic memory units and organizes them in an event-centric memory graph. Given a query, it first recalls relevant memory anchors through hybrid cues, then expands from these anchors along semantic and structural associations to recover missing supporting evidence. In this way, initially recalled memories serve not only as answer context, but also as cues for completing the evidence needed to answer. Experiments on LoCoMo and LongMemEval-S show that RippleMem achieves the best overall performance across evaluated settings, improving LLM-as-a-Judge accuracy by 3.95% on LoCoMo and up to 11.87% on LongMemEval-S, while reducing graph construction cost by about 30x.

• The paper introduces an evidence-conditioned memory system that combines typed event-graph associations with planned, anchor-guided expansion, achieving 87.14% LLM-as-a-Judge accuracy on LoCoMo and leading results on LongMemEval-S.
• RippleMem improves evidence-distributed recall by using semantic, structural, temporal, and participant cues, raising LoCoMo temporal-question accuracy by 9.66 judge-accuracy points and multi-session accuracy to 78.20%.
• The system offers a strong efficiency–accuracy trade-off by reducing offline construction time by roughly 30× and construction tokens by up to 69.3% versus LLM-mediated alternatives, while remaining limited to text-based benchmarks and one recollection round.

Motivation and problem formulation

RippleMem (2608.13334) addresses a specific failure mode in long-term memory for LLM agents: the information needed to answer a query is frequently distributed across temporally distant sessions, and one-shot query–memory matching often fails to assemble an answerable evidence set even when all required evidence has been stored. The authors frame this as an evidence recovery problem rather than a storage or retrieval-ranking problem. Flat retrieval may stop at the most directly matched record, while undirected graph expansion can traverse nearby but non-supporting memories. The paper's central design claim is that memory access should be evidence-conditioned: initially recalled memories should serve not only as answer context but also as cues for recovering missing support.

The design is informed by cognitive accounts of episodic memory—event segmentation, encoding specificity (Tulving), and pattern completion—but the authors are explicit that these are design intuitions rather than mechanistic claims about human memory.

Method

RippleMem follows a write–read architecture.

Write phase. Interaction history is processed as overlapping turn-level windows; a schema-guided LLM extractor converts each window into self-contained episodic memory units of the form $m_i = (r_i, \mathbf{v}_i, P_i, L_i, T_i)$, where $r_i$ is a canonicalized lossless restatement, $\mathbf{v}_i$ its dense embedding, and $(P_i, L_i, T_i)$ grounded participants, locations, and temporal intervals. Ungrounded cue fields are left empty rather than inferred. The memory bank is organized as a sparse weighted event graph with two typed edge channels: semantic association (cosine similarity over restatement embeddings) and structural association (Jaccard overlap over participants/locations plus exponentially decayed temporal-interval compatibility). Graph construction is incremental: each new unit scores a bounded candidate pool drawn from semantic nearest-neighbor search and cue-based indexes, retaining at most $K_{\mathrm{sem}}$ and $K_{\mathrm{str}}$ edges per node under fixed thresholds.

Read phase. Queries are decomposed into semantic, lexical, and structured cues that instantiate three complementary recall views whose union forms the initial evidence state $C_0$. A schema-constrained recollection controller then decides whether to continue, selects up to three anchors from the current evidence, and emits a missing-support target consisting of a target restatement and optional structured cues. Anchor-local expansion proceeds within two hops over both typed channels, candidates are scored against the target, and channel-specific results are merged by memory identity. The loop runs for at most one additional round ($R=1$) across all reported experiments. Evidence assembly deduplicates by memory identity and applies a deterministic source-aware ordering combining query alignment, retrieval provenance, and anchor status into a fixed top-$K$ budget ($K=30$).

A notable efficiency property follows directly from this design: persistent graph construction requires no LLM calls beyond memory extraction, since edge decisions come from deterministic similarity scoring.

Main results

On LoCoMo (1,540 questions), RippleMem achieves 52.49 F1, 44.05 BLEU-1, and 87.14 LLM-as-a-Judge accuracy, improving over the strongest per-metric baselines—SimpleMem on lexical metrics and RF-Mem on judge accuracy—by relative gains of 3.98%, 3.62%, and 3.95% respectively. On LongMemEval-S it reaches 84.80% under the SimpleMem-aligned setting and 86.60% under the EverMemOS-aligned setting, best overall in both comparison groups.

Gains concentrate on evidence-distributed question types, which supports the paper's causal account: on LoCoMo temporal questions RippleMem improves over SimpleMem by 5.45 F1 points and 9.66 judge-accuracy points, and multi-session accuracy rises to 78.20 versus SimpleMem's 60.92 and 80.45 versus EverMemOS's 73.68. On EverMemBench-Dynamic (2,400 questions over multi-party, multi-group histories), RippleMem attains the highest overall accuracy of 54.75%, a 4.44% relative improvement over the budget-matched RF-Mem baseline, indicating that the benefit comes from evidence-conditioned recollection rather than recollection-style retrieval alone.

Efficiency results are among the strongest claims in the paper. Compared with Mem0$r_i$0 and Zep, RippleMem reduces offline construction time by roughly 30× and construction tokens by 48.7–69.3×, while maintaining a compact 1,471.93-token average answer context. In an aligned comparison against MemPalace's documented Top-10 configuration, RippleMem achieves slightly higher judge accuracy (87.14 vs. 86.69) with approximately 5.1× fewer answer-context tokens.

Ablations and controlled comparisons

Three ablations isolate component contributions on LoCoMo. Removing anchor-local graph expansion causes the largest drop (87.14 → 83.12), confirming that first-hop recall alone misses stored support. Removing structural cues lowers accuracy to 83.83, consistent with structural associations recovering memories that share participants, locations, or time but differ in wording. Removing planned recollection drops accuracy to 84.35 despite identical expansion budgets, so the controller's anchor selection and target specification—not merely extra expansion—accounts for part of the gain.

An edge-construction replacement study shows that substituting GPT-4.1-mini edge decisions for typed association scoring slightly improves multi-hop judge accuracy (77.67 → 79.79) but reduces overall accuracy (86.62 vs. 87.14) while requiring ~27.7× more construction tokens and ~33× more calls. This is a useful negative result: flexible LLM-mediated edge judgment does not pay for itself at the system level.

On adversarial no-support questions (LoCoMo Category 5), RippleMem reaches 86.32% judge accuracy, matching SimpleMem but below the budget-matched RF-Mem (88.57%). The controller triggers a verification round on 93.0% of these adversarial cases versus 60.8% of answerable ones, reflecting a conservative response to insufficient first-hop evidence; removing the second round lowers Category-5 accuracy to 84.75%.

Limitations and open questions

The evaluation is confined to text-only conversational memory benchmarks; multimodal interaction, embodied agents, and tool-use environments are untested, and extending associative recollection to memories involving visual observations, actions, or tool states remains open. The pipeline depends on LLM-mediated extraction, query analysis, and recollection planning, which add latency and cost relative to single-pass retrieval—the recollection controller alone consumes an average of 2,880.6 tokens per question—and the paper concedes that end-to-end efficiency improvements remain necessary for large-scale deployment. Finally, current benchmarks provide limited stress tests for continuously growing personal memory over very long timelines; the effects of memory aging, evolving user states, and privacy-preserving deletion on associative recollection are unresolved. It should also be noted that several baseline scores on LongMemEval-S are taken from prior work rather than re-run, and the recollection controller's behavior was evaluated with a single additional round ($r_i$1), leaving deeper iterative recollection unexplored.

Conclusion

RippleMem reframes long-term memory access as adaptive associative recollection over a sparse event-centric graph, in which recalled memories act as cues for completing missing evidence. The combination of strong benchmark performance—particularly on temporal, multi-session, and open-domain questions—with a roughly 30× reduction in graph-construction cost indicates that deterministic typed associations plus bounded, planned expansion offer a favorable effectiveness–efficiency trade-off compared with both flat retrieval and LLM-mediated graph construction.