Episodic Memory
Version: 1.0.0 | Last updated: 2026-07-16
Purpose
Store attributable records of past events and decisions for later task-relevant recall.
Why
Raw conversation history is noisy and privacy-heavy; model-generated recollections can convert inference into false fact.
How
Write episodes only after a defined event boundary. Store event time, observed time, participants, tenant/ACL, source pointers, outcome, confidence, supersession, retention, and extractor version. Separate observed facts from model interpretations. Retrieve by authorized scope, temporal constraints, entities, and relevance; surface conflicts and source evidence. Correct or delete via lineage.
Tradeoffs
Curated episodes are more reliable than transcripts but may omit nuance. Preserve source pointers and allow abstention when evidence is incomplete.
Anti-patterns
- Saving every turn as permanent memory.
- Writing an agent’s summary as a verified fact.
- Retrieving stale episodes without effective-time filters.
Enterprise Considerations
Episodes often contain personal data. Enforce purpose limitation, subject access, retention, correction, and deletion across derivatives.
Checklist
- Fact and inference fields are separate.
- Event/effective times and provenance are retained.
- ACL, retention, conflict, correction, and deletion paths work.
- Recall usefulness and harmful-memory rates are evaluated.
References
Changelog
- 1.0.0 — 2026-07-16: Initial production standard.