Creativity · Agent Protocol
Agent Semantic Memory Pattern
Semantic memory is the agent's knowledge base of facts about the user, the world, and itself, abstracted away from the specific episodes that produced them. It's what lets the agent answer 'what do you know about me?' without replaying every past conversation. Typically stored as embeddings or a knowledge graph with deduplication and conflict resolution.
Protocol facts
- Sponsor
- Cognitive architecture research
- Status
- proposed
- Interop with
- Neo4j, Zep, mem0, vector databases
Frequently asked questions
How is semantic memory built?
Typically by extracting claims from conversation transcripts ('the user said their name is X') and writing them to a store, with deduplication against existing entries. Some systems use knowledge-graph triples (subject-predicate-object) for precise retrieval.
What if facts conflict?
Use a last-write-wins policy with audit trail, or a confidence-weighted merge. Critical: track provenance so you can answer 'why do you think my favorite color is blue?' and correct mistakes.
Vector store or knowledge graph?
Vector stores are simpler and work for most use cases. Graphs win when relationships matter ('who reports to whom', 'which file imports which') and when you need multi-hop reasoning. Hybrid systems (Zep, LightRAG) combine both.
Sources
- Zep — semantic memory architecture — accessed 2026-04-20
- mem0 — memory layer — accessed 2026-04-20