Keep your AI agent's personality and memory in files you own
An OpenClaw agent's character and memory are text files on its computer, so a model switch or shutdown can't take them. What changes, what to keep, how to move.
An OpenClaw agent’s personality and memory are plain text files on its own computer, not something held inside a model or on a provider’s servers. The model reads those files at the start of every conversation and writes new memories back to them. So a model can be switched, retired or cut off and the character and memory stay where they are: the next model reads the same files and carries on as the same agent, with its own way of saying things. Keep a copy of the agent’s workspace folder and you hold everything that makes it yours.
Where the character actually lives
OpenClaw keeps an agent in a workspace folder, ~/.openclaw/workspace by default, and its workspace docs list what each file is for. SOUL.md holds the persona, tone and boundaries; IDENTITY.md the name, vibe and emoji; USER.md what it knows about you and how you like to be spoken to; AGENTS.md its standing instructions. The memory is files too: MEMORY.md for lasting facts and decisions, and one dated file per day under memory/. The memory docs put it in one line: “The model only remembers what gets saved to disk; there is no hidden state.” What each file does, and how to edit them, is in the workspace files guide.
That is the difference from the memory a chat app keeps for you. A chat app’s memory lives in your account with the company that runs it, and it goes wherever that company takes it. An agent’s files are on the computer the agent lives on, in a format any text editor opens.
What changes when the model changes, and what doesn’t
Switching the model is one command or one message to the agent (both ways are here), and it touches none of these files. The name, the rules, the facts about you, every saved memory and the conversation history all stay on disk, and the next reply comes from the new model reading the same things.
What does change is the voice. The model is what reads the character and acts it out, so a different model reads the same SOUL.md differently: it may phrase things its own way, keep to the tone more or less closely, and pick up on a different share of what the files say. The agent is still yours and still knows what it knew; it may not sound exactly as it did. If a new model drifts, the fix is in the files, not the model: tell the agent what it got wrong and ask it to write that into SOUL.md, so every model after it reads the correction too.
What to keep so nothing is lost
The workspace folder carries the character and the memory. The rest of the agent lives beside it in ~/.openclaw/: its settings, its saved conversations, its connections to your messaging apps and the keys it uses. The workspace docs say those are never committed to the workspace’s repository, because they include credentials. So there are two copies worth having:
- The workspace, anywhere you like. OpenClaw makes the workspace a git repository when it creates it, and its docs recommend pushing it to a private repository as a backup. Ask your agent to do it and to keep it up to date; giving it a repo safely takes a key it makes for itself, no password pasted into a chat.
- The whole home, kept private. A copy of the full
~/.openclawfolder is what restores the agent exactly, conversations and connections included. Backing up and restoring covers doing that and checking the copy works.
The memory search index is not one of the things to worry about: it is built from the memory files and can be rebuilt from them on any machine.
Taking it somewhere else
Because the agent is files, moving it is copying them. The workspace docs describe the move: put the workspace on the new computer, point the agent’s configuration at it, and copy the saved conversations across if you want them too. The same files work whichever model the new setup uses.
With managed OpenClaw hosting from Everpod, the agent’s whole home is backed up every day, and if you ever want to take your agent with you, email us and we’ll send you a full copy of your pod’s data.