Why does agent memory go stale, and how do you catch it?
Your agent saved a fact from a spec. Then the spec changed. Here is why memory quietly goes stale, and how to catch it before an agent repeats it.
Jacob Cansler · Maker of Gryz
Updated 3 min read
Short answer: Agent memory goes stale when a fact is saved from a document and the document changes later. The fact keeps looking true, so every agent keeps repeating it. You catch it by keeping a link from each fact to the document and version it came from, and flagging the fact when that document changes.
How a fact goes stale
- An agent reads your launch plan and saves "the launch email goes out at 9am".
- A week later someone moves the email to noon and updates the plan.
- The saved fact still says 9am. Nothing about it looks wrong.
- Claude Code, Cursor and ChatGPT all recall it, and all three repeat 9am.
Most agent memory stores the fact and who saved it. Few keep where it came from, so there is nothing to check it against when the source moves on.
What it costs
- Wrong answers that sound sure. A stale fact comes back ranked high, because it was saved as important.
- Every agent repeats it. Shared memory is the point, so one stale fact spreads to every tool that reads it.
- Nobody knows to look. Without a link to the source, a person can't tell which facts to recheck after a doc changes.
How Gryz catches it
| Step | What happens |
|---|---|
| Save with a source | An agent saves the fact with the Gryz document it came from. Gryz checks the agent can read that document and that the quoted passage is really in it. |
| Keep the evidence | Gryz stores the version and the exact passage, so you can always see what the fact was based on. |
| Notice the change | When the document is edited, the fact reads as "source changed". When it is deleted, the fact reads as "source deleted" and keeps the passage. |
| Tell the agent | Recall and search show the change, with a note to re-read the document before relying on the fact. |
| Ask a person | The fact shows up under Needs you on your Memory page. Check it, then pick Still true, edit it, or forget it. |
You can also forget everything learned from one document in one step, and try a question on the Memory page to see exactly what an agent would get back.
This is part of what makes Gryz shared context rather than a pile of saved facts: memory, documents, tasks, skills and search that stay connected to each other.
FAQ
Does this work for facts that didn't come from a document? Facts said in chat have no document behind them, so there is nothing to watch. They still show who saved them and when.
Does Gryz rewrite a stale fact by itself? No. Gryz flags it and keeps the evidence. An agent or a person decides whether it still holds.
Does checking sources slow down editing documents? No. Gryz works out whether a source changed when the fact is read, so editing a document costs the same however many facts cite it.
Which agents can save facts with a source?
Any agent connected to Gryz over MCP, including Claude, Claude Code, ChatGPT, Cursor and Codex. They pass the document with the remember tool.
To try it, connect your agents, then see how saved facts work in the Agent Memory overview.
Related posts
The Four Kinds of Memory Your Agent Actually Needs
'Memory' sounds like one thing. It isn't. Here's how a fact, a preference, an episode and a procedure differ, and why your agent needs to know which is which.
Agent memory3 min
Gryz vs Mem0: which one should you use?
Looking for a Mem0 alternative? Mem0 is a memory API for apps you build. Gryz is shared context for your own agents: memory, docs, tasks, skills and search.
Comparisons5 min
Gryz vs ChatGPT memory: what's the difference?
ChatGPT memory is great inside ChatGPT. Gryz is shared context for every agent you use: memory, docs with comments and versions, tasks, skills and search.
Comparisons4 min