Blog

Why your AI agent keeps acting on a note that was never true

Aperisa ยท 2 min read

Checked on 2026-11-07.

One wrong line in an agent's memory file can send it in circles for a whole run. Anthropic's write-up on agent memory has a clean example: in a Pokemon Red run, Claude wrote that the PokeCenter was on the east side of Mt. Moon, declared the area done, and walked off a one-way ledge. The uncorrected run looped three times and never reached the next city.

If your agent gives confident answers that turn out wrong, the cause is often the same: it is reading a note it wrote earlier, not the source.

Why do agents act on wrong memory notes?

Anthropic describes four layers: transcripts, memory notes, memory representations such as wikis, and context injected into every session. The risk sits at the top. In their words, "an incorrect note can be carried into every session" once it reaches that layer.

Notes that describe a changing state go stale too. In the same article, Sonnet 3.7's detailed record of the game was "continuously going out of date".

What does this look like for a business agent?

Same pattern, less fun. For example, an agent writes "client X renews in March" after one email thread. Three weeks later the renewal moved. The note is still in memory, nobody checks it, and the next summary repeats it with full confidence.

How do you fix stale agent memory?

General fixes, for any setup:

  • Keep changing facts out of memory. Store where to find them, not the value.
  • Date every note and expire the ones about changing things.
  • Have the agent check a note against the source before it acts on it.
  • Review what reaches the always-loaded layer. That is the one place a wrong line costs the most.

Where does Aperisa fit?

Aperisa has read-only Google access (gmail.readonly, drive.readonly, calendar.readonly, contacts.readonly) and never sends, edits or deletes anything in your Google account. It keeps a searchable copy to answer from, so answers are only as fresh as that copy's last sync.

Give each agent its own key and set what it can reach: which data areas, work or personal, and how much detail, from full records down to totals only. Aperisa enforces those limits on the server. Revoke a key and it stops working within 2 minutes.

Access is by beta request at aperisa.com.

FAQ

Can an AI agent's memory be wrong? Yes. A note it wrote earlier can be false or go stale, and if it sits in the layer injected into every session, it is carried into each one.

Should an agent store changing facts in memory? Better to store where to find them, and read the value from the source when needed.

Which memory layer is riskiest? The one injected into every session, because a wrong line there touches everything.

Related: what controls what an agent can see and owning your data in the AI era. Background reading: Anthropic's agent memory write-up and the MCP spec.

Questions

Can an AI agent's memory be wrong?

Yes. A note it wrote earlier can be false or go stale, and if it sits in the layer injected into every session, it is carried into each one.

Should an agent store changing facts in memory?

Better to store where to find them, and read the value from the source when needed.

Which memory layer is riskiest?

The one injected into every session, because a wrong line there touches everything.

Published

Keep reading