RippleMem: From Retrieval to Recollection — Giving Agents Associative Memory
> *"Memory is not a folder in a drawer; it is a ripple on water — a stone drops, and the wave spreads across the whole pond."*
This post analyzes RippleMem, a memory architecture for LLM agents inspired by cue-dependent recollection from cognitive psychology.
Why flat retrieval fails at human-style recall
When asked "who did you have lunch with last Wednesday?", a human doesn't search a drawer. They start from anchors (the date, the location, the people) and spread associatively until a complete evidence chain is assembled. This is associative recollection.
Existing agent memory paradigms fall short:
- Full-context: stuffs all history into the context window — hits window limits and suffers "Lost in the Middle".
- Flat retrieval (RAG-style): returns top-k similar but *isolated* records, missing the connections between them.
- Graph memory: builds entity-relation graphs, but construction is expensive and compresses rich interactions into dry triples, losing time, place, and conversation content.
- LoCoMo: +3.95% LLM-as-a-Judge accuracy, with the biggest gains on evidence-dispersed questions.
- LongMemEval-S: +11.87% accuracy, strongest on multi-hop reasoning.
- Efficiency: graph construction ~30x cheaper than traditional graph memory — no full entity-relation graph is pre-built; spreading happens on demand at query time.
- Ablations: removing anchor spreading degrades performance significantly — the diffusion step, not hybrid retrieval alone, is the key driver.
- Granularity isomorphism: memory organization granularity should match the recall process granularity (cue → spread → assemble).
- Division of labor: let the LLM do semantic understanding; let the graph structure do association spreading.
- Externalized memory: offload memory *organization* into the graph so the LLM only understands, never stores.
- Evaluated only on text memory; no multimodal, embodied, or tool-use scenarios yet.
- LLM-mediated overhead: memory extraction, query analysis, and recall planning all use LLM calls, adding latency and cost.
- Existing benchmarks inadequately stress-test continuously growing personal memory.
RippleMem's architecture
1. Memory graph construction: each interaction is stored as a cue-rich episodic memory unit (full context preserved), linked by semantic and structural (time, place, person) edges into an event-centric memory graph. 2. Hybrid initial recall: combine semantic similarity + structural matching to find the most relevant memory anchors. 3. Anchor-local spreading: adaptively diffuse from anchors along semantic and structural associations — the "ripple" — pulling in related evidence as needed. 4. Evidence assembly: the activated memory fragments form the evidence set used as LLM context.
Key results
The cognitive-science analogy
RippleMem operationalizes Tulving's cue-dependent recollection (1972): human recall is not *search* but *reconstruction*. More cues ("a fruit" → "a red fruit") yield better recall. RippleMem's anchor-spreading mirrors this: start from cues, diffuse through an association network, assemble the full memory.
It also resonates with other principles:
Three design principles for agent memory
1. Memory is organization, not storage — well-organized memories are recallable; poorly organized ones are isolated records. 2. Retrieval is spreading, not locating — pull in related evidence from an anchor along an association network. 3. Context is preservation, not compression — keep full episodic units; add associations at the organization layer without compressing content.
Limitations and future work
Conclusion: the paradigm shift from retrieval to recollection
RippleMem's real contribution is a paradigm shift: agent memory should be designed around recollection (cue → spread → assemble) rather than retrieval (locate → return). Once memory moves from retrieval to recollection, an agent stops being "a database with search" and becomes "a memory system that associates."
Paper: arXiv:2608.13334
Code: no official repository provided, but the method is reproducible with LangChain + Neo4j + an LLM API.