When to reach for it
- The agent needs to learn from past cases.
- Recurring tasks follow similar solution paths.
- Context including time, goal, and outcome is relevant.
Episodic Memory.
Past tasks, persisted, so the agent can recognise its own déjà vu.
Completed interactions are stored as discrete, timestamped episodes — a natural-language memory stream — and retrieved later by a combined recency, importance, and relevance score, so the agent can reuse what worked in a similar past situation.
In practiceA software-debugging agent stores each resolved bug as an episode, then retrieves the three closest past episodes by embedding similarity when it encounters a new error.
When it backfires
- Past cases quickly become obsolete.
- Storing personal data is problematic.
- A static semantic knowledge base suffices.
The tradeoff
Experience-based adaptation is gained against high curation and data privacy maintenance efforts.
What it actually does.
Past episodes are stored and retrieved by similarity.
Two ways this pattern will hurt you.
Stale episodes mislead the planner
Episodes from a deprecated tool surface stay matchable. The planner picks an old strategy and runs it against an API that no longer exists.
Fix · Tag episodes with a tool-surface version. Filter or expire episodes when the surface they referenced changes.
Similarity alone retrieves the wrong episode
Ranking by embedding similarity only, a superficially-worded match outranks the genuinely relevant recent success, and the agent replays a strategy that never applied here.
Fix · Score retrieval by recency + importance + relevance combined (per Generative Agents), not similarity alone, and weight successful outcomes above mere textual match.
Threat exposure
What this pattern exposes
Adopting this pattern opens these attack surfaces. Each links to its entry in the threat model.
Full threat model →Where Episodic Memory is native.
Microsoft Agent Frameworkepisodic storeNative
Keep going