Episodic, Semantic, and Procedural Memory: What Should an Agent Store?

Episodic memory represents an event, semantic memory represents a fact or concept, and procedural memory represents a way to perform a task. For AI applications, these are useful design categories rather than a requirement to build three separate databases.
The categories help decide what evidence to retain and how much authority to give it. “A retry worked once” is an event. “This endpoint rejects expired tokens” is a supported fact. “Refresh the token before retrying” is a procedure that still needs validation.
Start with the question the record should answer
A support agent may need to answer what happened in the last ticket, which account the user currently has, or how to handle an approved refund. Each question needs a different kind of record.
| Type | Question | Important metadata |
|---|---|---|
| Episodic | What happened? | Time, participants, source event |
| Semantic | What is currently known? | Scope, provenance, validity |
| Procedural | What should be done? | Version, approval, applicability |
One source can support several records, but do not promote them automatically. A customer quote can describe an event without establishing a general product fact. A successful workaround can remain an observation until it has been tested and approved.
Keep procedures under explicit control
Procedures influence future actions, so they need stronger controls than conversational notes. An instruction embedded in a retrieved document must not redefine tool permissions. A remembered workaround must not override a current runbook because it was recently mentioned.
Store the procedure's applicability conditions and version. When a workflow changes, remove or supersede the old instruction from the paths that supply current behavior. The original incident can remain available for historical explanation.
Choose retrieval by task
For an incident recap, retrieve relevant events and their sources. For an account question, consult authoritative current state. For a workflow, retrieve the approved procedure and confirm that its conditions apply.
The categories can share storage while using different filters and ranking rules. A single similarity search over all three can overemphasize a vivid old incident when the user needs a current rule. The memory architecture guide describes how these policies fit into the larger application.
Test transitions between categories
Use an event that suggests a possible fact but does not prove it. Check that the agent preserves the uncertainty. Next provide a confirmed source, then an approved procedure based on that source. Finally, revise the procedure while leaving the original event intact.
Inspect whether the agent can explain what happened without recommending obsolete steps. That is a stronger test than asking it to repeat definitions of memory types.
To explore a managed implementation, start with Supermemory and a small set of labeled records. Preserve source references and approval boundaries in your application, and verify the result with the lifecycle checklist.