Agent memory
Agent memory is information an AI agent keeps from earlier tasks or conversations, such as preferences, past decisions and facts about a party, so it can use them later.
What it means
A language model starts each task knowing nothing about your business. Memory is the stored notes that are fed back in when the agent works: a customer who always wants delivery before the 5th, or a supplier whose invoices carry freight separately. Short-term memory covers the current task; long-term memory persists across days. Anything remembered should be visible to staff and correctable.
Example
A garment exporter's agent notes that a buyer in Surat rejects quotes without packing charges shown separately. Three weeks later it drafts a new quote with packing as its own line, without being reminded.
Records and checks
- Record each remembered item with its source, date and who it concerns.
- Verify that old or wrong notes are removed, and that one customer's notes never appear in another's work.
- The process owner reviews remembered items periodically and approves deletions.
Common mistakes
- Letting the agent save guesses as facts, so a one-off exception becomes a permanent rule.
- Never expiring old notes, so outdated prices or contacts keep being used.