Creativity · Agent Protocol
Agent Episodic Memory Pattern
Episodic memory is the 'what happened' layer of agent memory: timestamped, narrative, specific. It complements semantic memory (general facts) and procedural memory (learned skills). For long-running assistants, episodic memory is how the agent answers questions like 'what did we decide last Tuesday?' or 'why did you configure the deploy that way?'
Protocol facts
- Sponsor
- Cognitive architecture research
- Status
- proposed
- Interop with
- mem0, Zep, LangGraph memory, vector databases
Frequently asked questions
Episodic vs. semantic memory?
Episodic = 'on 2026-01-15 Adarsh told me his Firebase project is called anvevoice'. Semantic = 'the user's Firebase project is named anvevoice'. Agents usually need both; episodic preserves provenance and lets you reconstruct reasoning.
How is episodic memory stored?
Typically as event records with timestamp, actor, action, and embeddings for retrieval. Retrieval combines recency, relevance (vector similarity), and importance scores — the Generative Agents paper is the canonical reference.
How do I prevent episodic memory from exploding?
Apply consolidation: periodically summarize clusters of related episodes into higher-level semantic facts, and age-out raw episodes past a TTL. Otherwise retrieval latency and cost grow unboundedly.
Sources
- Generative Agents paper — accessed 2026-04-20
- MemGPT paper — accessed 2026-04-20