// docs
Memory
S.E.N.T.R.I gets smarter the longer you work with it. What it learns about you and your estate is kept between conversations, filed under the work folder it belongs to, and grows into a neuron you can open and manage from the chat.
What it remembers
Memory is for what the agent could not work out again by looking: how you like to work, the corrections you have given it and why, how your estate or codebase is wired, the decisions you have made and the reasons behind them, and where things live.
It deliberately does not store anything a command would tell it — host lists, job status, file contents — because those change, and a stale copy is worse than none. And it never stores secrets.
| Type | Holds |
|---|---|
| user | Who you are and how you work: your role, your preferences, how you want answers. |
| project | How this estate or codebase is set up, and the decisions behind it. |
| feedback | Corrections you have given, with the reason, so the same mistake is not made twice. |
| reference | Where a thing lives: a URL, a path, a ticket. |
How the agent uses it
- An index to start from — every conversation begins with the list of memories, a name and a one-line hook each. The hook says what the memory tells, not just what it is about, so the agent knows when to open it.
- Read before relying — a memory's full text only reaches the agent when it opens it, and it opens it before acting on the detail rather than guessing from the hook.
- Update, don't duplicate — before saving, it checks the index. If a memory already covers the subject, that one is rewritten instead of a near-copy being added.
- What it sees wins — a memory records what was true when it was written. If it contradicts what the agent can see now, the agent trusts the present, tells you, and corrects the memory.
- Linked — related memories point at each other, so one fact leads to the next.
One memory per work folder
Every memory is linked to a work folder, so each project keeps its own and nothing from one leaks into another. Change the chat's working folder and the memory changes with it; two agents in the Agent Workspace on different folders each work from their own.
Whichever provider you run — Claude Code, Codex CLI, OpenRouter or Ollama — reads the same memory, so switching model never means starting from nothing.
The neuron
A work folder's memory is drawn as a 3D neuron that grows with every conversation. Open it from the chat to see what the agent knows about the current work folder, and manage it right there.
Managing memories
The agent saves a memory on its own when it learns something worth keeping, and you steer it in plain language:
- “Remember that our domain controller is DC01” — saved as a new memory.
- “That's wrong, it moved to DC02” — the same memory is rewritten, not duplicated.
- “Forget the note about the old build server” — that memory is removed.
Memory is not the chat history
Keep secrets out of it