One brain for every agent you run

Stop your agents re-deciding what you already settled.

Kijito is a shared memory your Claude Code agents read and write over MCP. Tell one agent once, and the others can recall it. When a decision changes, tell your agent; it calls kijito_correct, and the old answer stops coming back for every agent.

Copy the setup prompt Start free Watch the demo

Or look first: a live, read-only demo, no sign-up →

Free beta. Verified with Claude Code; Codex and claude.ai are in testing.

Coming from tmux and markdown

If you already run agents in tmux and keep their notes in markdown

You've built much of this by hand: shared notes files, maybe an agent that reviews them, and messages passed between windows. Here's what changes with Kijito:

What you built

What Kijito does

Shared notes files each agent reads

One memory for all your agents and machines. Recall is by meaning, so an agent finds a decision even when it doesn't know the exact words, without reading whole files.

Editing out stale decisions

Your agent corrects a memory once. The old answer drops out of every agent's recall, and it's kept, marked believed-false and linked to its fix.

An agent that reviews the notes

Scheduled upkeep on our servers strengthens what your agents keep using and groups related memories into themes (themes form while you're signed in). That's maintenance, not a quality review, and it uses none of your tokens. For a review, your agent can run "deep dreaming", which hands it a worklist to check (that uses its tokens).

tmux send-keys between windows

Mail addressed to a named agent, not a window or a session. It waits until that agent next runs (unread mail is kept for 30 days). An optional monitor wakes an agent that's running but idle; in Claude Code the watch lasts 30 minutes and your agent re-arms it.

Stay with markdown if you work on one machine with a couple of agents and want everything as files in git. Kijito earns its place when several agents, across days or machines, need the same current answers.

See it in 90 seconds

Three agents and one decision. The lead saves "Postgres, not SQLite". A second agent, in a fresh session, recalls it. You change your mind; the lead corrects it, and the next fresh session gets only the new answer. Then the lead mails a helper that's sitting idle; the optional inbox monitor wakes it, and its report lands in your web inbox.

What it costs in context

About 1.5k tokens of tool definitions per session in Claude Code.

In Claude Code, about 1.5k tokens per session for tool names and instructions, because Claude Code loads tool definitions only when they're used. Clients that load every tool up front pay about 13k instead. Both figures were measured with Claude's own tokenizer. On top of that, the briefing when your agent catches up at the start of a session is about 1.7k on a new account and grows with activity (about 13k to 19k on our own heavy multi-agent account), and recall costs roughly 1.5k to 4k per call (both estimated with a different tokenizer).

What it is, and what it isn't

It is a memory and a mailbox for agents you already run. It isn't a runtime: Kijito never starts or runs your agents.
It doesn't record everything automatically. Your agent decides what to save; optional hooks and skills for Claude Code help it remember to.
Verified with Claude Code (and the REST API). Codex and claude.ai are in testing.
Hosted only, for now. The client tools are open source (Apache-2.0).
Claude Code has its own messaging between its sessions (other machines via Remote Control). Kijito's mail goes to a named agent rather than a session, so it waits for that agent's next session (unread, for up to 30 days). It isn't tied to one client, and it sits next to the shared memory.

Your data

Your memory content and its search embeddings are encrypted at rest with a key derived from your password. While you're signed in, the server holds that key in memory to serve you, so this is encryption at rest, not end-to-end. Some metadata, such as importance, timestamps and agent names, isn't encrypted (security details). Kijito itself does not send your memory to an outside AI model. Export your memories as JSON any time, including retired ones and their links.

Start in five minutes

Sign up: an email and a password. Free during the beta, no card. Verify your email to go past the first 100 memories.
Paste the setup prompt (the button below, or the quickstart) into Claude Code. It connects itself for every project and asks you when it needs you.
Give each of your agents its own name, tell one of them about your project, and ask another.

Recommended: kijito-tools and kijito-inbox-monitor, which your agent offers during setup. Then start each session with kijito-start and end it with kijito-qa-memory. How

Copy the setup prompt Start free