research(memory): A-MEM Zettelkasten-style evolving link weights for SYNAPSE graph

[bug-ops]

Source

A-MEM: Agentic Memory for LLM Agents (arXiv:2502.12110, NeurIPS 2025)
https://arxiv.org/abs/2502.12110

Summary

A-MEM structures episodic memories as Zettelkasten-style notes: each stored memory gets structured attributes (keywords, tags, contextual description) plus explicit links to related memories. Links are continuously strengthened, weakened, or removed as new memories arrive — the graph evolves organically rather than being built once and queried statically.

Applicability to Zeph

Relevance: HIGH. Zeph's SYNAPSE already does spreading activation over a static graph. A-MEM's link-weight evolution (based on co-retrieval frequency and temporal decay) would make the SYNAPSE graph self-organizing over time without requiring manual curation.

Key differences from current implementation:

• Current: edges have fixed confidence scores, never updated after insertion
• A-MEM: edge weights evolve based on co-retrieval frequency and recency
• A-MEM: each memory note has structured attributes for richer retrieval

Implementation sketch

• Add retrieval_count and last_retrieved_at columns to graph_edges
• On each SYNAPSE recall pass, increment retrieval_count for traversed edges
• Periodically decay weights of un-retrieved edges (multiplicative decay)
• Expose edge weight as a factor in BFS node scoring (higher-weight edges = higher activation)

Complexity: MEDIUM

Schema changes are additive. Weight update logic can be integrated into the existing spreading activation pass.