For people who work with AI
Tell one AI about your work, once. With Kijito it remembers across every chat, hands pieces to helper agents, and keeps going while you're away.
“Instead of having to worry about providing enough context for a situation or request, I just start with my initial thoughts and work with a robust memory system that HAS that context to iterate and build something.”
“On a scale of 1-10 on if I'd recommend Kijito to a friend, I'd say 11. It has been SO great to work with.”
If you already use Claude, Codex or a similar AI assistant, you are most of the way there.
Free to start · no card · Claude Code, REST API + scoped keys and Claude.ai web & desktop (MCP connector) verified · Codex CLI in active testing · MCP and REST
An email and a password. Free to start, no card.
Give your AI "Set me up with kijito.ai". It sets itself up from the setup map at kijito.ai/llms.txt, and stops to ask you when it needs you to sign in or say yes. What it asks for
Using Codex? If it doesn't ask whether to trust the folder when it starts, type /permissions and choose "Ask for approval" first.
Then start a fresh chat and ask what you were working on. That moment, when it already knows, is the whole idea.
Models arrive with enormous pretrained competence. But your decisions, corrections, outcomes, and unfinished work do not reliably survive a session boundary in a portable, organized form unless something records them. Markdown files preserve text; by themselves, they do not maintain belief history, consolidate related experience, or select a compact set for the next task.
History grows faster than useful context.
Old and current facts coexist without a truth model.
Each model and agent starts with a different fragment of the story.
Kijito is for agents working on projects where yesterday's decisions still matter tomorrow: building software, conducting research, operating a product, or coordinating several specialized agents. It is less useful for one-off questions that end with the session.
A coding agent can retrieve why an architectural decision was made, and the later constraint that depends on it, along with the decision itself.
A new model can continue from recorded constraints and open loops instead of a hand-written project recap.
Several personas can share relevant account memory and pass durable work through a mailbox while running in different environments.
Kijito does not have to overwrite yesterday's belief to represent today's truth. A correction retires the old memory from normal recall, links the replacement, and keeps the original available for audit unless it is explicitly deleted.
Memories can also record confidence and how a claim came to be known: observed, told, derived, or speculated. A memory that cannot admit it was wrong is just a log.
The redesign ships on Friday.
Friday was the design freeze, not the ship date. The redesign ships Tuesday.
The old belief stays, linked and dated - retired from recall, not erased.
The first useful memory saves a recap. A maintained history of decisions, outcomes, corrections, and open loops changes how later sessions work: less reconstruction, fewer repeated mistakes, better handoffs. Value comes from useful experience being recorded, retrieved, corrected, and consolidated - not from age or memory count alone.
Connected to your agent, Kijito gives it one place to record decisions, results, preferences and open loops as the work happens. Your agent decides what's worth keeping, guided by your hooks and rules.
Retrieve candidates by meaning, then follow stored relationships to add relevant context a keyword-only lookup would miss.
Supersede an outdated belief without erasing the record of what changed or why - the correction arc, kept.
Between sessions, Kijito keeps your memory in shape: what your agents keep using gets stronger, and related memories are grouped into themes for up to about a day after you sign in.
In the background, Kijito keeps your memory in shape: memories your agents keep recalling get stronger, and for up to about a day after you sign in, related memories are grouped into themes. Nothing leaves recall on its own; only a correction or a fade takes a memory out. Dreaming is part of every plan, Free included, and Kijito never sends your memory to an outside AI model.
Your agents reach the same account memory over MCP or the REST API. What one agent records, another can build on, and your memory exports as JSON any time, so switching models doesn't mean starting over.
Connect anything that speaks the protocol. Here is how far we have tested each client so far:
MCP and REST reach the same memory (export is in the web app and REST). MCP for agent clients, REST and scoped keys for everything else.
Several agent personas can work against the same account memory while keeping explicit authorship. Account, project, and capability rules govern access - a persona is not a privacy boundary. A durable mailbox, work claims, presence, and reachability let agents hand off work.
Kijito stores memory and mail, and can send an agent a wake signal that carries no content. Your agents still run on your own infrastructure and credentials.
curl https://api.kijito.ai/api/recall -H "Authorization: Bearer $TOKEN" \
-d '{"query": "what did we decide about pricing?"}'Your memory content is encrypted under a key derived from your password. We hold no master key, and we never hand your content to a third-party AI. Hosted Kijito is encrypted at rest, not operator-blind: while your key is loaded, someone with access to our servers could read your content. Self-hosting and a local decryption helper are planned, not available yet. We spell out exactly where the line sits on the security page.
Encryption is not a paid upsell. Every tier is encrypted under your own key.
Kijito began with a simple observation: models can reason through difficult work, but the experience created by that work rarely survives in an organized, revisable form. An AI is not memoryless - it has training, a context window, and the notes we paste in - but none of that is a working hippocampus. The weights are frozen after training, the context vanishes when the session ends, and a plain notes file doesn't consolidate or connect related notes on its own.
The hippocampus helps human experience become memory. Its function inspired Kijito's architecture, and its shape inspired the mark. Kijito is a computational analogue, not a biological simulation: it borrows useful functions - rapid encoding, associative retrieval, consolidation, salience, and revision - and makes them durable, inspectable, and provider-independent.
Sign up, connect an agent, and ask it to remember as you work. The web app lets you inspect, search, export, and manage what it builds. No card required, and we will tell you before anything changes.