Memory types
Working, episodic, semantic, procedural — four kinds of memory, and the rule that only what is retrieved back into the context window can influence a decision.
The context window is an agent's working memory, and it is both small and temporary. Anything the agent should remember beyond the current task has to live somewhere else — and, crucially, be brought back into the window when it is needed. "Memory" in an agent is not one system but a set of stores with different contents and different retrieval rules.
Start here
Only what is in the context window can affect a decision. External memory is not memory until it is retrieved. So the design question is never just "what should we store?" but "what will we pull back in, and when?" — retrieval is the memory, in practice.
One task — which memories can the agent reach?
recalled into context (3)
Working memory: the current task: follow up with this customer
Episodic: last week they reported a duplicate charge
Semantic: refunds are allowed within 30 days
Working memory is small and always in play; the other stores live outside the context window and only matter once they are *retrieved back in*. An agent is only as good as what it pulls into context at the moment it decides.
Toggle the external stores and see what the agent can actually recall for one task.
The four kinds
- Working — the live context: the current task, recent steps, the plan. Fast, always present, and gone when the session ends.
- Episodic — what happened, and when: past conversations, past tool results, past mistakes. "Last week this customer reported a duplicate charge."
- Semantic — facts, policies, and knowledge, independent of when you learned them: "refunds are allowed within 30 days."
- Procedural — how to do things: skills, tool recipes, routines.
"To schedule a follow-up, call
create_calendar_invite."
Inside vs outside the window
Working memory is the only place computation happens; the rest are storage. Bringing an item in costs context (the budget lesson from the foundations chapter), so retrieval must be selective. Too little and the agent forgets what it needs; too much and the signal drowns.
Careful
Memory is a liability as well as an asset. A stale episodic memory ("this customer is on the free plan") is worse than no memory, because the agent trusts it. Stored personal data is a privacy and compliance surface. Treat every store as something that must be corrected, scoped, and expired — not just written.
Check yourself
Eduspheria wiki · Agentic AI, Memory & context
0 / 4 answered
Next: the retrieval that turns stored knowledge into usable context.