Memory & State

'Agent memory' is three problems wearing one name:

• Working memory — the context window itself: this run, this task, this budget. Gone when the run ends.
• Episodic memory — past runs: 'last time the user asked this, we tried X and it failed.' Retrieve by similarity, inject as a note.
• Semantic/procedural memory — distilled facts and learned playbooks: 'user prefers terse output', 'deploy requires the staging flag'. Structured, curated, long-lived.

Conflating them is the classic mistake: dumping chat history into a vector store gives you episodic noise, not semantic knowledge.