UnisonMemory types
Episodic, semantic, and procedural memory for AI agents - what each type stores, when to use which, and how Unison classifies them at search time.
Agent memory comes in three types, and Unison lets you filter recall by each (memoryType on search: episodic, semantic, procedural, or auto).
| Type | Stores | Example |
|---|---|---|
| Episodic | what happened - events, sessions, decisions in time | "On June 3 we migrated the queue to pgmq" |
| Semantic | what is true - facts, entities, relationships | "The worker service depends on Postgres 17" |
| Procedural | how to do things - runbooks, conventions, skills | "Deploys go through the staging smoke suite first" |
Why the distinction matters
A query like "what did we try for rate limiting?" wants episodic memory - the history, including failures. "What is our rate limit?" wants the current semantic fact, not the trail. "How do I roll back?" wants procedure. Collapsing all three into one vector soup is why naive memory systems return the wrong kind of right answer.
In Unison, episodic material accumulates as documents and timeline events; semantic truth lives as bitemporal facts that supersede each other; procedural knowledge is documents that stay current by edit, not append. Default auto blends them; pass memoryType when your agent knows what kind of answer it needs:
unison search "rate limiting" --memory-type episodicWhat is AI Memory?
AI memory is the layer that lets agents retain and recall knowledge across sessions - distinct from the context window, distinct from RAG. Types, architecture, and how teams share it.
Agent Memory vs RAG: What's the Difference?
RAG retrieves documents you indexed; agent memory accumulates what your agents learned while working. Different questions - most teams need both.