# Episodic memory for AI agents

> A reference on episodic memory in AI agents: memory for timestamped events rather than facts, the four stores (working, episodic, semantic, procedural), seven requirements an episodic store has to meet, the patterns that work, how the benchmarks (LoCoMo, LongMemEval) test it, and what is still unsolved. Version 1.0, 3 November 2026.

## Full text

- [llms-full.txt](https://episodicmemory.ai/llms-full.txt): the entire reference in one fetch.

## Key points

- Episodic memory stores events with their context and time; semantic memory stores facts stripped of occasion. They are written, indexed and retrieved differently.
- RAG reads a corpus someone else wrote. Episodic memory writes a record of what the agent did and saw, then reads it back. The write path is the part RAG does not have.
- An episodic store needs: timestamps separate from write time, context binding, a cheap write path and selective reads, scheduled consolidation, decay, provenance from fact back to episode, and deletion that reaches embeddings and backups.
- The event log is the source of truth; vector indexes and graphs are derived views that must be rebuildable.
- Retrieval needs two paths: similarity for "what is like this", temporal range for "what happened between these points".
- Context stuffing is not memory. Compare memory systems at matched token budgets, or the long-context baseline wins for the wrong reason.
- Memory is an attack surface: anything written into long-term memory is read back later, so a record must be treated as untrusted input.

## Benchmarks

- LoCoMo (arXiv 2402.17753) — very long-term conversational memory, questions asked many sessions after the events.
- LongMemEval (arXiv 2410.10813) — five abilities: information extraction, multi-session reasoning, temporal reasoning, knowledge updates, abstention.

## Citation

Episodic memory for AI agents, version 1.0, 3 November 2026. https://episodicmemory.ai/