One memory.
Any shell.
Your agent learns a project as it works. Hermit keeps what it learns in one place, so the next model, the next tool or the next laptop starts where the last one stopped.
What gets
remembered.
Lasting things, saved by the agent as it works. You don't have to ask. Each memory is a line or two, and remembers which model wrote it.
Facts
staging is read-onlyHow things are in this project.
Decisions
small PRs, no new depsA choice, and the reason for it, so the next model doesn't undo it.
File map
api/stripe — webhook handlersA key path and what it is for. The path and a line, never the file.
Preferences
pnpm, not npmHow you like things done: tools, tests, commits, tone.
Big model, everything.
Small model, the digest.
One memory, packed to fit the context window. The agent loads it at the start of a session; what it gets depends on the room it has.
The whole map
While a project's memory is small, or when the model asks for it, every line comes back: facts, decisions, files and preferences, grouped by kind.
The lines that matter
Once it grows: pinned lines first, then the newest of each kind, about 1,500 tokens. Set it lower for a small local model. The rest is a keyword search away.
Yours, and
only yours.
- Never source codeHermit can't see your disk or repo. It stores what the agent sends: short notes, never files. Secrets become [redacted] before anything is written.
- Encrypted at restMemory text is encrypted with AES-256-GCM and bound to your account. It is decrypted only to answer your own requests.
- Export or deleteTake a project as JSON or Markdown, from the app or the API. Forget one memory or a whole project, anytime.