Memory Hygiene for Long-Running AI Assistants

Memory makes Rocky more useful, but only when it stores the right information. Bad memory creates stale assumptions, privacy risk, and confusing behavior.

Good things to remember

• User preferences that improve future work
• Stable project names and public URLs
• Repeated workflow rules
• Approved style guidelines
• Non-secret environment notes
• Lessons learned from verified fixes

Things not to remember

• Passwords, API keys, tokens, cookies, or recovery codes
• Unnecessary personal data
• Private customer details that are not needed later
• Temporary facts that will quickly become stale
• Speculative conclusions not verified against a source

Memory maintenance

Memory should be updated when a preference changes, a project moves, a provider changes, or a workflow is retired. If Rocky keeps making the same wrong assumption, the memory is stale or too vague.

Profiles and boundaries

Use profiles for serious separation. Do not depend on memory alone to isolate client, personal, and production work.

Verification

For current facts, inspect the live source when available. Memory is context, not proof. A repository, website, dashboard, or API should override remembered assumptions.