lontar — one memory for all your AI agents

How it works

Connect your agents. One URL and a sign-in — no database, no environment variables, nothing to install per machine.
They save as you work. Your agents record decisions, corrections and preferences under rules you can read, and save whenever you ask.
Your memory is there at the start.
- Claude Code loads your memory at the start of a session, automatically.
- For ChatGPT, Codex, VS Code with Copilot and Gemini CLI, lontar tells the agent to fetch your memory when a session starts.
- Everywhere else, you ask once.
Within 90 days of launch, the same automatic start for the two most-requested agents that support session-start context injection.

Why lontar

Capture happens in your agent, under rules you can read. lontar's built-in skill has your agent save decisions, corrected assumptions and preferences — reading before writing, asking when unsure — and saves whenever you ask. Nothing is extracted in the background.
Will be open source, never trained on, yours to take or delete. lontar will be open source (AGPL-3.0). Your memory is never used for model training, or shared for it, on any plan. Export all of it at any time. Delete anything and it is gone from the live database at once, and from backups within 8 days.
Walls between contexts. Work memory and personal memory stay apart. lontar's tools can't name a space, so a model can't ask for another one.
Corrections keep history. When something stops being true, lontar records the correction instead of overwriting it. Ask what was true on a date, or what you believed on a date.

Works with

Claude Code · claude.ai web, Desktop and mobile · Cursor · ChatGPT · Codex · VS Code with Copilot · Gemini CLI · OpenCode · Goose · Hermes Agent · OpenClaw · Antigravity — and any MCP agent, and a CLI for the rest.

Price

$10/month or $100/year. 14-day free trial, no card. One plan, everything included.

Questions

What happens if it isn't built?

If 280 people confirm, hosted lontar gets built. If the window closes short it may be extended once, and the date on this page is always the current one. If it ends short, everyone who confirmed gets an email saying so, the list is deleted 30 days later, and lontar stays a personal tool. You are not putting your name down for something that quietly never arrives.

Is there a free plan?

No free plan. There is a 14-day free trial with no card, and then $10/month or $100/year. One plan, everything included.

Is my memory used for training?

No. Memory content is never used for model training, or shared for it, on any plan. There is no cheaper plan that trades training rights for a discount, and lontar has no model of its own to train.

Can I self-host it?

lontar will be open source (AGPL-3.0), so you are free to run your own copy. It will not be supported: no setup documentation, no support, and no promise that your copy stays compatible. What a subscription pays for is the running service — reach across agents, one memory on every machine, backups, upgrades and patches.

How is this different from the memory my agent already has?

Built-in memory lives inside one vendor's surfaces and stays there. What you tell one agent is invisible to the next, and it does not come with you when you switch tools or machines. lontar is one memory that every agent you use reads and writes, under rules you can read — and you can export all of it, or delete it, at any time.

What does joining share?

Your email address, and — only if you fill them in — which agents you use and whether you are happy to talk for 15 minutes. We also record which link brought you here. There are no cookies and no free-text fields. Visits are counted with a hash that changes daily and cannot be linked across days, and nothing identifying is stored. You confirm by clicking a button in an email; every email carries a one-click unsubscribe that deletes your address; and an address that is never confirmed is deleted after 14 days. The privacy notice lists all of it.