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

RippleMem is a long-term memory system for LLM-based agents that shifts memory access from isolated retrieval to adaptive associative recollection. The paper argues that the main bottleneck in long-horizon agent memory 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.

The paper identifies a central failure mode: "memory access can still fail to assemble an answerable evidence set. Flat query-centered lookup may stop at the most directly matched record, while graph expansion without an explicit evidence need may traverse nearby but non-supporting memories. The key question is: when an initially retrieved memory is relevant but incomplete, how should memory access continue to recover the missing support needed for answering?"

RippleMem's design is informed by cognitive accounts that organize experience into event-like units and view recall as cue-dependent and associative. It operationalizes these ideas "through a write–read design: the write phase stores interaction history as cue-rich episodic memory units linked by semantic and structural associations, while the read phase uses initially recalled memories as cues for recovering additional support."

Write Phase: The write phase converts dialogue history into self-contained event memories with semantic representations and episodic cues and links these memories through semantic and structural associations. Each memory unit follows a fixed schema: mi = (ri, vi, Pi, Li, Ti), where "ri is a canonicalized event restatement that can be interpreted independently of the original dialogue window, vi is the dense representation of ri, and Pi, Li, and Ti denote grounded participants, locations, and temporal cues or intervals when available. The extractor is instructed to resolve speaker-dependent references into explicit participant names, anchor relative time expressions to absolute intervals only when supported by the dialogue timestamp."

The Event-Centric Memory Graph is a sparse weighted event graph G = (M, E) with E = Esem ∪ Estr containing typed links retained after association scoring. Semantic association captures meaning-level proximity between memory restatements using cosine similarity of dense vectors. Structural association captures shared grounded episodic cues using Jaccard overlaps over participants and locations, and temporal compatibility based on interval overlap or exponential decay for nearby intervals. Graph construction is incremental and sparse, retaining up to Ksem semantic links and Kstr structural links per node above score thresholds.

Read Phase: The read phase uses this event memory space to recover evidence for a query. It begins with Hybrid initial recall that extracts retrieval cues cq from a query q, including semantic, lexical, and grounded episodic cues, instantiating three complementary recall views: C0 = Csem ∪ Clex ∪ Ccue.

Then Memory-anchor planning invokes a schema-constrained recollection controller Πrec that outputs whether to continue (dr), the selected anchors (Ar ⊆ C), the missing-support target (gr), and an optional stop reason (sr). The controller decides whether further recollection is needed and selects anchors that are central to the query or likely to evoke missing support through graph associations.

Anchor-local expansion then expands only from the selected anchors within a bounded graph neighborhood: Ur = m ∈ M C dG(m, Ar) ≤ h. Candidates are scored against the missing-support target: the target restatement is used for semantic matching, and the target cues are used for structural matching. The top candidates from the two channels are merged by memory identity to form the round-level support set Sr.

Evidence Assembly: The final evidence state is consolidated by memory-unit identity and ordered using a deterministic source-aware ordering with score ρ(m) = λq a(q, m) + λp π(m) + λa Ianc(m), combining query–memory alignment, retrieval provenance, and anchor status. The final evidence context EK consists of the top-K memory units under this ordering.

Main Results: On LoCoMo, RippleMem achieves 52.49% F1, 44.05% BLEU-1, and 87.14% LLM-as-a-Judge accuracy, with relative improvements of 3.98% and 3.62% in F1 and BLEU-1 over SimpleMem, and judge accuracy over RF-Mem by 3.95%. On LongMemEval-S, RippleMem achieves the best overall accuracy in both comparison groups, reaching 84.80% under the SimpleMem evaluation setting and 86.60% under the EverMemOS evaluation setting.

The paper highlights Larger gains on evidence-distributed questions: on LoCoMo temporal questions, RippleMem improves over SimpleMem by 5.45 F1 points and 9.66 judge-accuracy points, and on LongMemEval-S it improves multi-session reasoning over SimpleMem (78.20 vs. 60.92) and EverMemOS (80.45 vs. 73.68).

Ablations: Removing graph expansion causes the largest drop in judge accuracy, from 87.14 to 83.12, indicating that first-hop recall often misses supporting evidence. Removing structural cues lowers accuracy to 83.83, and removing planned recollection drops to 84.35.

Efficiency: RippleMem reduces construction time by approximately 30× and construction-stage tokens by 48.7–69.3× compared with Mem0g and Zep, while maintaining a relatively compact average answer-context budget of 1,471.93 tokens per question.

Additional evaluations: On EverMemBench, RippleMem achieves the highest overall accuracy of 54.75%, representing relative improvements of 28.13% over MemOS and 4.44% over the budget-matched RF-Mem baseline, performing best on seven of the nine subtasks. On adversarial no-support questions (LoCoMo Category 5), RippleMem achieves 86.32% judge accuracy, with the additional recollection round improving from 84.75%.

Limitations: The paper notes RippleMem is evaluated primarily on text-only long-term conversational memory benchmarks and does not cover multimodal interaction, embodied agents, or tool-use environments. It also uses LLM-mediated operations for memory extraction, query analysis, and recollection planning, which add latency and cost compared with single-pass retrieval baselines. Finally, current benchmarks provide limited stress tests for continuously growing personal memory over very long timelines.

Conclusion: The paper concludes that reliable long-term memory depends not only on storing past experience, but also on organizing it so that partial recall can guide further evidence recovery, and that recollection-aware memory access offers a promising direction for building more reliable, efficient, and context-sensitive LLM agents.

Improvements to AI Systems Based on RippleMem:

1. Implement adaptive, multi-hop memory recollection instead of single-pass retrieval. The AI system will iteratively assess whether its current evidence set is sufficient to answer a query; if not, it will select specific memory anchors (events, participants, locations) and expand only within a bounded graph neighborhood to recover missing supporting evidence, rather than stopping at the first relevant match.

2. Add event-centric memory structuring with typed associations. The system will store past interactions as self-contained episodic units (with canonicalized restatements, participants, locations, temporal cues) and link them via both semantic similarity (dense vector proximity) and structural overlap (shared entities and temporal compatibility). This enables the system to recover evidence that is contextually related but not lexically similar to the query.

3. Introduce cue-dependent recall with hybrid retrieval views. The system will extract three complementary cue types from a query—semantic, lexical, and grounded episodic cues (participants, locations, time)—and use them jointly to initialize recall, improving recall for questions that reference specific entities or timeframes rather than only topical keywords.

4. Enable schema-constrained recollection planning. The system will use a controller that explicitly decides whether to continue recollection, which anchors to expand from, and what missing-support target to seek. This prevents over-expansion into irrelevant memories and focuses computational effort on evidence that is likely to resolve the query’s incompleteness.

5. Implement source-aware evidence ordering and assembly. The system will rank final retrieved memory units by combining query–memory alignment, retrieval provenance (e.g., initial recall vs. expanded), and anchor status, ensuring that the most reliable and directly relevant evidence is prioritized in the final context passed to the generation model.

6. Add robustness against no-support queries. The system will be able to recognize when a query has no answerable evidence in memory and explicitly stop recollection, avoiding hallucinated answers. This is achieved by the planned recollection controller’s stop mechanism, which improves accuracy on adversarial or unanswerable questions.

7. Optimize memory construction for long-horizon agents. The system will build its memory graph incrementally and sparsely (retaining only top-K semantic and structural links per node), reducing construction time by 30× and token usage by up to 69× compared to dense graph methods, while maintaining high retrieval quality—enabling deployment in resource-constrained or real-time agent settings.

8. Improve multi-session and temporal reasoning. The system will handle questions that require evidence distributed across multiple past interactions or across time, by using temporal cues and structural associations to link events that occurred at different times but share participants or locations, leading to higher accuracy on multi-session reasoning tasks.

9. Enable budget-controlled evidence retrieval. The system will maintain a compact answer-context budget (e.g., 1,500 tokens per question) by limiting expansion depth and using planned recollection, allowing it to fit within LLM context windows without sacrificing answer completeness.

10. Provide a generalizable memory access layer for LLM agents. The system can be integrated as a drop-in memory module for any LLM-based agent (e.g., conversational assistants, task planners, research agents) to improve long-horizon task performance, reduce hallucination from incomplete memory, and handle complex queries that require assembling evidence from multiple past events.