Most agent memory stores conflate what happened with what's true. The conflation is why the memory degrades — and the separation is the fix.
The cognitive science distinction maps cleanly onto agents: episodic memory is the events (what happened, when), and semantic memory is the facts (what's true). An agent that stores everything as episodes drowns; an agent that stores everything as facts loses context. Both stores, separately, is the design.
First Principles: What Each Is For
- **Episodic.** What happened, in order. Used for recent context, recurrence detection, and as raw material for consolidation.
- **Semantic.** What's true, across episodes. Used for decisions, preferences, and facts the agent acts on.
Conflating them means the agent either can't find the forest (all episodes) or can't find the trees (all facts, no context).
The Design: Two Stores
The design is two stores with a bridge: an episodic store for events, a semantic store for facts, and a consolidation step that moves lessons from one to the other. The agent retrieves from each by need — recent context from episodic, decisions from semantic.
- Episodic store: events, ordered, time-bounded.
- Semantic store: facts, structured, queryable.
- Consolidation: episodic → semantic, on a cadence.
What Breaks When Conflated
- An agent that remembers every event as a fact and acts on one-offs.
- An agent that can't recall what just happened because it only stores facts.
- A single store that grows until retrieval is noise.
How to Use Each at Query Time
At query time, the agent retrieves recent episodic context (what happened in this conversation, what happened recently) and relevant semantic memory (what's true about this user, this task). The two together are what makes the agent feel like it remembers — because it does, in the right way.
Conclusion
Episodic and semantic memory are different stores for different jobs. Keep them separate, bridge them with consolidation, and retrieve from each by need — and the agent remembers in a way that compounds instead of degrades.
About FACTA
FACTA helps startups and growth-stage teams turn AI into production systems that keep running — not demos that impress once.
We design the architecture around the parts that actually break under real usage: tooling you own, credentials you control, failover, cost controls, observability. The boring infrastructure that keeps a system alive after launch.
Led by Matías Baglieri and Carolina Fogliato, we focus on one thing:
AI leadership that builds. Not just advises.
Ask us whether your agent's memory separates episodes from facts.
We'll tell you what conflating them costs. See memory consolidation for the bridge.
Explore AI Automation