AI agent memory: sources, freshness, and deletion rules

AI agent memory is the information an assistant retains or retrieves across tasks. Reliable memory design separates temporary conversation context, durable preferences, authoritative business records, and audit logs. Store only justified information, verify current facts before consequential actions, and provide a clear way to correct or delete retained data.

Separate memory by purpose

Session context helps the current conversation and can expire when no longer needed. Durable preferences may support later interactions if their retention is appropriate. Retrieved documents provide evidence, while business systems hold authoritative records such as current estimates or job status. Tool state tracks identifiers and progress. Audit logs serve investigation and accountability. Keeping these purposes separate prevents a casual note or old summary from acquiring the authority of an approved contract.

Define what may become durable

Write explicit rules for permitted facts, approval requirements, and prohibited content. A confirmed communication preference may be useful; a speculative complaint summary should not become a permanent customer label. Exclude secrets and unnecessary sensitive information. Record the source, date, purpose, and owner of each durable item. Treat an inference differently from a customer-confirmed fact. If the workflow cannot explain why a piece of information will be useful later, reconsider whether it should be retained at all.

Make freshness part of retrieval

Use timestamps, effective dates, expiry, and supersession rules appropriate to the information. Customer preferences may change, supplier prices can become stale, and appointment availability needs a live check. A retrieved memory about an old estimate must not override the current approved proposal. When sources conflict, prefer the documented authority or escalate the question. Do not let semantic similarity alone determine which record governs a decision involving money, scheduling, or customer commitments.

Support correction and deletion end to end

Let authorized people inspect and correct retained facts. Identify copies in caches, indexes, summaries, and backups when designing deletion behavior and retention limits. Explain operational limitations honestly instead of promising immediate removal from every storage layer without evidence. Track which applications consume the information so corrections propagate. NIST’s AI Risk Management Framework provides a useful structure for considering risks, responsibilities, and ongoing monitoring across this data lifecycle.

Test whether memory improves the task

Create cases with expired preferences, replaced proposals, conflicting sources, deleted records, and attempts to store prohibited information. Inspect what was retrieved and whether it influenced the final action appropriately. Compare performance with and without memory to determine whether retention adds value. Measure stale-fact use, correction success, unsupported inferences, and task completion. Begin with a narrow memory purpose and expand cautiously. More retained text is not inherently better; useful memory is relevant, current, authorized, and accountable.