AI engineering
Conversation Memory With Clear Scope and User Control
Useful AI memory needs a purpose, an owner, and an expiry policy. Separate working context from lasting preferences and keep stored facts easy to correct.
Conversation history is not automatically useful memory. A long transcript may contain outdated decisions, temporary requests, sensitive details, and statements that were never verified. Anthropic’s context-engineering guidance discusses managing the information available to an agent. For an application, the design question is which information should persist, for whose benefit, and under what controls.
Separate three kinds of context
Working context supports the current task and can often disappear when the task ends. Conversation history records what was said. Durable memory stores selected information intended to influence later sessions. Give these categories different retention, access, and correction rules rather than treating one transcript as all three.
Store durable entries with their source, owner, creation time, and scope. A personal preference should not become an organization-wide instruction. A project decision should not silently apply to another project. When a fact is inferred instead of explicitly provided, represent that uncertainty and consider whether it belongs in persistent memory at all.
Design correction before recall
Imagine a hypothetical design assistant that remembers a customer prefers concise project summaries. That preference may be useful across sessions. A temporary instruction to omit budget discussion from one draft should not become a permanent rule for every future report.
Give users a simple way to view, edit, and delete saved preferences. When new information conflicts with an old entry, use an explicit update policy rather than appending contradictory memories forever. For business facts, retrieve current records from the source system when possible. Memory should not replace a live account status, inventory count, or permission check.
Evaluate memory as a source of errors
Test cross-project contamination, stale preferences, shared-device scenarios, and a user correcting something remembered incorrectly. Check whether deleted entries remain in summaries, search indexes, or cached responses. A visible delete control is incomplete if derivative stores continue influencing later answers.
Keep recalled material bounded and relevant to the current task. Record which memory entries were used so a poor response can be investigated without exposing the full history to every operator. Treat remembered text as data, especially when it originated in external content. A stored instruction from an untrusted document should not acquire authority merely because it was saved earlier.
