Long-Term Memory Patterns: Profiles, Preferences, and Facts in Agentic AI
Long-Term Memory Patterns: Profiles, Preferences, and Facts
What belongs in long-term memory
- User profile basics (name, role, product plan)
- Preferences (tone, format, units)
- Stable facts (project name, recurring context)
What should NOT be stored
- Passwords, OTPs, private identifiers
- One-time transactional data (unless required)
- Anything user didn’t consent to
Schema matters
Store memory in typed fields, not one blob. Example structure:
- preferences.format = “HTML”
- preferences.language = “en”
- work.project = “Edugators Courses”
Typed memory is easier to validate, update, and delete.
Keeping memory truthful
Long-term memory can become outdated. Use:
- timestamps
- confidence scores
- user-confirmation prompts when uncertain
