Does your AI actually use its memory?
Connecting a memory is not the same as the AI using it. We measured how often an AI agent saves and looks things up, and what changes the answer.
The short answerNot always. In our hardest Claude Code test, a few lines in its instructions file took saves from 17 of 24 to 23 of 23. Rewording the memory tool's own description made things worse.
What we tested
We run an automated test that drives real Claude Code sessions with Contexel connected, gives each one a request such as "remember that we ship on Thursdays" or "save to memory: ...", and records whether the save reached Contexel, went only to the agent's own built-in memory, or went nowhere. It also checks that questions about the team get looked up and that unrelated requests are left alone.
To make it realistic, we set up the hardest case a customer might hit: the agent gets no instructions from the memory server, the connection has a random name, and the agent already has a lived-in memory of its own.
What we found
| Setup | Saves that reached the memory |
|---|---|
| Hardest case, nothing added | 17 of 24 |
Hardest case, plus a short rules block in CLAUDE.md | 23 of 23 (one run errored before calling any tool) |
| Hardest case, with "save it to memory" added to the tool's description | 12 of 24 |
| A fresh install, instructions delivered, no rules block | 10 of 10 saves, 5 of 5 lookups |
Three lessons
- What loads at the start decides it.An AI app's own memory is part of its built-in instructions. A connected memory has to be named just as early, in a rules file or in the server's own instructions, to compete.
- More words about "memory" can backfire.Adding "or save it to memory" to the save tool's description pushed the agent toward its built-in memory instead, and saves dropped from 17 to 12 of 24.
- Connection speed matters.When the memory server connected too slowly, its instructions arrived after the session had started and were never read. Once we made it connect faster, a fresh install saved 10 of 10.
What this means for you
Whatever memory you connect, give your AI one line telling it to save there and check there. In Contexel, the console's Connect tab writes it for you: a rules block for Claude Code and Cursor, a line for GitHub Copilot or Gemini CLI, and one sentence to paste into ChatGPT's or Claude's instructions. It's optional, and it makes saving and lookups more reliable.
These tests ran in Claude Code. The Claude and ChatGPT apps behave differently, and we check them by hand.
Questions
Does this apply to ChatGPT and Claude on the web?
The numbers are from Claude Code. The lesson carries over: an app that is told about the memory at the start of a chat uses it more reliably, which is what the one-sentence instruction does.
Does the AI stop using its own memory?
No. The instruction asks it to also save to Contexel, never to skip its own memory. In our tests the agent often kept a copy in both.
When the AI missed, where did the memory go?
Into the agent's own built-in memory file, or nowhere. In every miss, the agent never called a Contexel tool.