# Kijito > Kijito is a provider-independent memory and continuity layer for AI agents. It stores durable experience, retrieves by meaning and association, preserves belief history through corrections, consolidates memory over time, and supports durable multi-agent handoffs over MCP and REST. For a growing class of useful work, model intelligence is no longer the primary bottleneck. Continuity is. Kijito gives the agents you already use one durable place to record decisions, outcomes, corrections, and unfinished work — then retrieve, revise, and hand off that experience across later sessions and supported clients. It is an external memory layer; it does not replace your model or agent runtime, and it does not change model weights. This file is a concise, descriptive index for developers and agents. Every claim below is also stated on one of the linked pages. ## What it does - Recall by meaning and association — a compact set of relevant experience, without reloading the whole history. - Corrections supersede an outdated belief while keeping the original available for audit, so memory can admit it was wrong. - Scheduled maintenance strengthens what your agents keep using and groups related memories into themes; nothing is removed from recall unless your agent corrects or fades it. - The memory lives outside any one model provider, so verified and compatible clients can reach the same evolving record — and you can export it. - Several agent personas can share the same account memory and hand off work through a durable mailbox, while Kijito routes opaque wake signals — your agents still run on your own infrastructure and credentials. ## Trust boundary Hosted Kijito encrypts memory content and embeddings at rest under user-derived key material, decrypting transiently in RAM to answer an authenticated request. Encrypted at rest — not operator-blind. Operator-blind decryption runs off our servers: via Enterprise on-premises self-hosting, or a local helper you run — both on our roadmap. ## Metaphor Kijito is a prosthetic hippocampus for AI: the hippocampus helps turn what happens into memory and experience, and that function inspired Kijito's architecture and the shape of its mark. It is a computational analogue, not a biological simulation. The neuroscience background lives on /docs/concepts. ## Set up (for agents) If a person asks you to set them up with Kijito, this is the whole path. The person does four things: create the account, open the sign-in link and sign in again, answer the setup questions, and approve each setup action. Before you start: the inbox monitor (step 7) installs with uv (https://docs.astral.sh/uv/). If `uv --version` fails, tell the person first; it installs with `curl -LsSf https://astral.sh/uv/install.sh | sh` or `brew install uv`. Where pipx is already installed, `pipx install kijito-inbox-monitor` works instead. kijito-tools (step 6) runs with npx, which comes with Node.js. In Codex: if your network requests fail and your approval policy is granular with sandbox_approval false (escalations are rejected without asking the person), Kijito is not down and you cannot fix it from inside. Tell the person to type /permissions and choose "Ask for approval", or to restart Codex in their project folder and choose "Trust and continue" when it asks "Trust this folder?". Then retry. 1. Account (the person): they sign up at https://kijito.ai/signup and keep their recovery codes. 2. Connect: add the MCP server https://api.kijito.ai/mcp for every project, not just the current one. The exact command for Claude Code, Codex, OpenCode and Claude.ai is on https://kijito.ai/docs/quickstart. 3. Sign in (the person, in their own client): the person starts the sign-in themselves. In Claude Code that is /mcp, then kijito, or `claude mcp login kijito` in their own terminal. In Codex it is `codex mcp login kijito` in their own terminal. Never run the sign-in command from your own shell to pass the link on: the link only works while the command that printed it is still running, and when your command exits the person's sign-in ends in "localhost refused to connect". Before they start it, give them the two sign-in warnings. These words get lost when paraphrased, so this one line is the exception to saying things in your own words: Say to them, exactly: "A sign-in link prints in your terminal and is not opened for you — cmd-click it, or copy it into your browser. You may be asked to sign in again even though you are already signed in — that is the OAuth consent step, not an error." In Codex, they then start a new Codex session so the Kijito tools load. 4. Onboard: after the first kijito_startup, run kijito_onboard(action="start"). This is the setup, not an optional extra. It gives you one step at a time, with the words to say: the person's one-sentence authorization for what setup will do, a persona to file your work under, then global instructions, tools and the inbox. Relay each step, act only on a yes, record the answer with kijito_onboard(action="answer", step=..., answer=...), and call start again for the next step. From then on pass that persona on every write; without it, memories are filed under "personal". 5. Make it global: GET https://api.kijito.ai/api/setup-text?client= returns the short standing-instructions block and where it goes for that client. Add it additively; never remove what is already there. After the person says it is in place, read the file back before you tell them it is done. 6. Tools (recommended; suggest them to the person and install only with their yes): https://github.com/KijitoAI/kijito-tools adds the kijito-start / kijito-qa-memory skills, and in Claude Code also session catch-up and context tracking. Claude Code: npx kijito-tools (then restart Claude Code; hooks and skills load at session start). Codex: npx kijito-tools --provider codex --skills-only (the two skills go to ~/.codex/skills). The habit to suggest, once installed: start every session with kijito-start, and end it (or clear it) with kijito-qa-memory. In Claude Code that is /kijito-start and /kijito-qa-memory; in Codex, $kijito-start and $kijito-qa-memory. 7. Inbox monitor: https://github.com/KijitoAI/kijito-inbox-monitor wakes a running agent when mail arrives. Install: uv tool install kijito-inbox-monitor (or pipx install kijito-inbox-monitor). With kijito-tools 0.2.11 or later (step 6; an older install can miss a dead monitor and never wake you, so re-run npx kijito-tools@latest), start your inbox and prove it works in one command: ~/.claude/kijito-inbox-start.sh --persona . You are done only when it exits 0; exit 3 means only your agent's side is left (it prints the line to arm), and exit 2 means something is missing (it prints the fix). Without kijito-tools, do it by hand. Check it can reach your inbox: kijito-inbox-monitor --persona --self-test; this proves the token and the connection, not the wake. Then run it for real and keep it running (README: https://github.com/KijitoAI/kijito-inbox-monitor#running-the-producer-for-real-supervision), and connect it to your agent so each event wakes it (README: https://github.com/KijitoAI/kijito-inbox-monitor#waking-your-agent). Use the persona from step 4. Prove it: send yourself one message with kijito_hive_send and confirm it woke you. Installed is not the same as working. 8. Confirm: call kijito_startup, remember one thing, and recall it in a fresh session. ## Canonical pages - [Home](https://kijito.ai/): what Kijito is and who it is for. - [Live demo](https://kijito.ai/demo): a read-only Kijito account with a staged example (memories, one corrected decision, agent mail); the same data as JSON, no credentials, at https://api.kijito.ai/api/demo - [Quickstart](https://kijito.ai/docs/quickstart): sign up, connect Claude Code, Claude.ai, Codex, or OpenCode with one command, and complete a first remember/recall. - [Connect a client](https://kijito.ai/docs): the connection guide for other MCP clients and the REST API. - [Concepts](https://kijito.ai/docs/concepts): events versus experience, recall, corrections and belief history, consolidation, forgetting, and the hippocampus analogy. - [Security and trust](https://kijito.ai/docs/security): the encryption model and exactly where the hosted trust boundary sits. - [Connections](https://kijito.ai/docs/connections): connecting each client, with per-client verification status. - [Hives and multi-agent](https://kijito.ai/docs/hives): shared account memory, durable mailbox, work claims, presence, and wake routing. - [API reference](https://kijito.ai/docs/api): the REST API and scoped keys. - [REST index](https://api.kijito.ai/api/index): machine-readable JSON list of every REST route (method, path, auth, scopes, parameters); no credential needed. - [FAQ and troubleshooting](https://kijito.ai/docs/faq): connection, token expiry, recall behaviour, and account limits. - [Pricing](https://kijito.ai/pricing): Free during the beta — up to 10,000 memories. Paid billing is off during the beta. We will give advance notice before enabling it, and no account is charged without an explicit checkout. - [Status](https://kijito.ai/status): live service status and help. - [Changelog](https://kijito.ai/changelog): what has shipped. - [Does a corrected memory stay corrected?](https://kijito.ai/blog/does-a-corrected-memory-stay-corrected): a published revocation attack's test cases, run against Kijito's correct, fade and update. - [Terms](https://kijito.ai/terms), [Privacy](https://kijito.ai/privacy), [DPA](https://kijito.ai/dpa), [Acceptable Use](https://kijito.ai/acceptable-use): legal.