Source

A-MEM: Agentic Memory for LLM Agents (NeurIPS 2025)

Finding

A-MEM treats each memory write as a mini-agent action: when a new note is stored, an agent generates structured attributes (contextual description, keywords, tags) and dynamically links the note to related existing notes via embedding similarity and keyword overlap. Memory organization is itself agentic — the graph structure evolves with each write rather than being built from a fixed extraction schema.

Applicability to Zeph

Zeph's graph memory extracts entities/edges from message content via LLM (chat_typed). A-MEM's write-time linking step is complementary: after entity extraction, a second fire-and-forget pass computes similarity between the new memory record and existing ones, creating direct edges between related memories (not just entity co-occurrence). This improves multi-hop recall for queries that span multiple sessions.

Implementation Sketch

1. After spawn_graph_extraction writes entities+edges, spawn a follow-up link_memory_note task
2. For each new entity, query embedding store for top-K similar entities (threshold 0.85)
3. For each match above threshold, upsert an edge (new_entity) -[SIMILAR_TO]-> (existing_entity)
4. Tag the edge with the similarity score and timestamp

This reuses the existing embedding infrastructure (zeph-memory EmbeddingStore) and graph edge schema with no new schema required.

Priority

Medium — improves multi-hop recall for long-running agents. Token cost is one embedding lookup per new entity (cheap).