Hot, Warm, and Cold Agent Memory: Tier by Use and Freshness

Tier agent memory according to how quickly information is needed, how often it is reused, and what it costs to retrieve. “Hot,” “warm,” and “cold” are design labels rather than universal product features. A useful tiering policy also specifies promotion, invalidation, permissions, and what happens when slower evidence is unavailable.

Define the tiers in application terms

Hot context might include the current task and a small set of confirmed preferences. Warm memory might be searchable decisions and prior interactions. Cold storage might retain source documents or transcripts that are accessed only when a question needs their detail.

Do not assign fixed latency numbers to these labels without measuring the actual system. A remotely stored “hot” profile can still be slow, and an indexed historical document can be quick to retrieve. The labels should describe policy and expected use, not substitute for instrumentation.

Promote information because it is useful

Repeatedly used, stable information can be promoted into a compact representation. Keep its source reference and validity state. A summary of a contract is useful only if the application can detect that the contract changed.

Avoid promoting a fact simply because the model repeated it. Repetition can amplify an early error. Use confirmed utility, explicit user preference, or a tested rule to decide what belongs in always-available context.

Invalidate across every copy

Corrections and permission changes must reach cached profiles, summaries, indexes, and retained sources according to the product's lifecycle rules. A corrected warm record does not help if the prompt still includes a stale hot copy.

Test this sequence: store a preference, promote it, correct it, then ask from a new session. Repeat after removing access. Record which representation supplied the answer and whether the old value remains eligible.

Budget for slow-path retrieval

Define what the agent does when cold evidence is needed: wait, answer with an explicit limitation, or continue through a background task. A short prompt should not force unsupported certainty when detailed evidence exists elsewhere.

The context-eviction guide explains which retrieved items to omit from a prompt. To evaluate retrieval and memory together, try Supermemory with questions that alternate between recurring preferences and rarely used source details; measure both paths before choosing a tiering policy.