Giving an AI agent memory is easy. Keeping it true is the hard part.

Description and use

AI agent memory is easy to persist, but its truthfulness decays quickly. Adding re‑check commands, separating observation from inference, and external verification prevent decisions based on stale facts.

Problem

AI agents can store memory across sessions, but the stored facts can become stale or incorrect, leading the agent to act on false information.

Outcome

By adding re-checks, separating observation from inference, and verifying externally, the agent’s decisions become more reliable and stale memory is quickly identified and corrected.

Workflow

• 1. Store each memory fact as a markdown file with a clear claim.
• 2. Add a re-check command that can be executed to verify the claim.
• 3. Separate observed evidence from inferred conclusions in the file.
• 4. Prioritize filesystem state over memory when they conflict.
• 5. Perform external verification of claims (e.g., public API checks).
• 6. If external check fails, investigate the checker before updating memory.
• 7. Regularly review and update memory entries that influence decisions.

Evidence

In a single working day last week, four separate memory entries turned out to be confidently wrong

Source

Write the re-check, not just the claim

Source

Let files outrank memory

Source