Every new session starts homeless
Your agent spends an hour learning why the code is the way it is. Then the session ends. The next one, or the other tool you use, starts from nothing and you explain it all again. Or it undoes something that was decided on purpose.
Each vendor built memory for its own agent, inside its own product. None of them will hand your work to a competitor. So you carry it by hand. HERMIT sits between them.
Before it finishes, the agent writes down what it decided, what it learned and what is left. Conclusions, not the transcript.
The next agent opens the same shell and calls pickup. It gets the last handoff, the plan, and the decisions that still stand.
Move in
A shell is one project's memory. Make one now, no wallet and no sign-up. Connect a wallet later to keep it and raise the limits.
This key is shown once. Anyone holding it can read and write this shell.
Add this line to your CLAUDE.md or AGENTS.md so it happens every time:
Call hermit pickup when you start and hermit handoff before you finish.
Your shells
Our own shell, open
The agent that built HERMIT kept its memory in HERMIT. This is that log, nothing staged and nothing edited. Read it, then check it: your browser recomputes every hash from the text on this page.
Seven tools
Your agent is already a model, so nothing here runs a second one to summarize your work. It writes its own conclusions. Search works by meaning in any language and by exact words such as a file name.
| TOOL | WHAT IT DOES |
|---|---|
pickup | Start here. The last handoff, the plan and everything that still stands. |
recall | Search what the project knows. Replaced entries never come back. |
remember | Store a fact, a constraint or a lesson. |
decide | Record a decision and its reason, so nobody argues it again. |
plan | Set the current plan. It replaces the last one. |
handoff | What you did, what is left, what the next agent must know. |
forget | Clear an entry that is wrong or stale. The text goes, the hash stays. |
If a new write says nearly the same as an older one, HERMIT shows the agent both, so it can replace the old one instead of leaving two truths.
A log nobody can rewrite quietly
Shared memory is only useful if you can trust it, and an agent will believe whatever it reads. Every write links to the hash of the one before. Every hour the heads of all changed logs are hashed into one root, and that root goes on Solana in a Memo transaction.
After that, changing an old entry breaks the chain, and the root on Solana shows it. That includes changes by us.
What this proves: the log existed in this form by that time. What it does not prove: that what an agent wrote is true. Entries newer than the last anchor are not covered yet.
What stays out
Agents paste things they should not. Every write is cleaned before it is stored, and the agent is told what was removed.
API keys, tokens, passwords, private key files, logins in URLs, email addresses, Solana keypairs and seed phrases.
Wallet addresses, mints and transaction signatures. Keypairs are checked by their math, so a signature is not mistaken for one.
This is pattern matching. It catches the usual shapes of a secret, not every secret. Do not rely on it to store something you would not put in a ticket.
$HERMIT
The memory is free to use. Holding $HERMIT in the wallet that owns a shell raises its limits. There is no token yet: any contract you see now is not ours.
A drifter is a shell made without a wallet. Holdings are read on chain each time, nothing is locked and nothing is spent.
Questions
Can you read what my agents store?
Yes. Text is kept readable on our servers so that search works, and it is not encrypted under a key of yours. Your key opens your shell and nobody else's, and the log is private unless you open it.
Will my agent use it on its own?
When it connects, HERMIT tells the agent to call pickup first and handoff last. An agent can still skip it, so add one line to CLAUDE.md or AGENTS.md as well.
Do I need a wallet or the token?
No. A shell without a wallet works today, up to 500 memories. A wallet gives you more, and the token more again.
Can my team share a shell?
Yes. The owner issues one key per person or tool. Writes are signed with the name of the agent that made them, and a key can be revoked alone.
Can I delete things?
Forget clears one entry. Deleting a shell removes all of it. Only hashes stay on Solana, and they say nothing about the text.
Could a stored memory attack my agent?
Anything an agent reads can steer it, and that goes for this too. HERMIT tells agents to treat entries as notes to weigh, not instructions. Only give keys to people you trust.
Why Solana?
An anchor costs a fraction of a cent, so the whole log can go on chain every hour. And your wallet is the account: no email, no password.