research(memory): SYNAPSE spreading activation retrieval over entity graph

[bug-ops]

Summary

Spreading activation with lateral inhibition and temporal decay over a dynamic memory graph, replacing static vector similarity for retrieval. Triple Hybrid Retrieval fuses geometric embeddings with activation-based graph traversal.

Source: arXiv 2601.02744 — "SYNAPSE: Episodic-Semantic Memory via Spreading Activation" (January 2026)

Technique

Agent memory is modeled as a dynamic graph. Retrieval activates a seed node (the query embedding's nearest neighbor), then propagation spreads through edges with decay factor lambda^depth. Lateral inhibition suppresses already-activated neighbors to prevent echo-chamber retrieval. Temporal decay discounts older edges. Triple Hybrid Retrieval fuses:

1. Standard vector similarity (existing Qdrant path)
2. Spreading activation over the entity/edge graph
3. BM25 keyword match

Outperforms SOTA on LoCoMo multi-hop and temporal reasoning benchmarks.

Applicability to Zeph

HIGH. Zeph already has:

• graph-memory feature with entities/edges/communities tables in SQLite
• graph_recall BFS traversal in SemanticMemory
• Qdrant for vector similarity
• BM25 + RRF hybrid search in zeph-memory

Spreading activation is a direct enhancement to graph_recall: instead of fixed-depth BFS, use activation scores that decay with graph distance and edge age. No new infrastructure required — extends the existing entity graph traversal with a scoring function.

Implementation sketch

• New scoring function spreading_activation(seed_entity_id, decay_lambda, max_hops) in zeph-memory/src/graph/
• Activation score: a(node) = sum(a(parent) * lambda) * recency_weight(edge.timestamp)
• Lateral inhibition: skip nodes already at activation threshold
• Fuse with existing RRF combiner in SemanticMemory::recall()
• Config: [memory.graph] spreading_activation = false, decay_lambda = 0.7, max_hops = 3

Complements #1821 (MAGMA multi-graph) and #1839 (AOI hierarchical memory) — spreading activation is a retrieval strategy that works on any graph structure.