Your agents forget.
fractal remembers for them.
Every fact in its place, in a tree that organizes itself; updated when it changes; separate for each agent. Connects in two minutes to Claude, ChatGPT, Cursor and any MCP client.
claude mcp add --transport http fractal https://getfractal.it/mcp
Claude Code — in your terminal
Works with Claude · ChatGPT · Cursor · VS Code · Codex · Gemini · LangChain · any MCP client
Every number is re-measured by an automated test on every code change.
A memory that does the work itself.
Corrects itself, never contradicts
When a fact changes, the new one takes the old one’s place. The old one stays a month as a copy, then goes away on its own.
Knows what was true on a day
Every memory has two dates: when it was true and when you learned it. Ask how things stood on March 3, and it knows.
Organizes itself
No folders or tags to pick: every new memory finds its topic, and a topic that grows splits on its own.
Skills written once
Write a procedure in one place: Claude, ChatGPT, Cursor and your agents always get its latest version.
Finding similar text is not remembering.
A memory knows what changed, when it held, and where it belongs.
| What happens | A vector database alone | fractal |
|---|---|---|
| A fact changes | Both versions stay, and the agent may pick the old one | The new one takes the old one’s place, which stays in the history |
| «What was true on March 3?» | It can’t tell: it keeps no time | You ask with memory_asof |
| 20,000 memories | Compares the question with all of them, or with an index to tune | Goes down by topic: 24 ms |
| A wrong memory | You delete it, and tomorrow the agent infers it again | You dismiss it: gone from results, kept on record as rejected |
Every agent already has its own memory.
Create an agent and its memory is there, separate from the others’. If one should read another’s, you allow it: one arrow in the dashboard, and every exchange is logged.
Ready for the enterprise.
One memory per department, permissions decided by a person, and nothing leaves Europe. The guarantees live in the code: every point below is on today.
Security
Memories encrypted on disk and in backups (AES-256). We keep only a fingerprint of API keys. Two-step verification, mandatory for the whole company if you decide so.
Control
One memory per department. An agent reads another’s only with a permission, which can be revoked, and every exchange is logged.
Access
OAuth 2.1 for ChatGPT, VS Code and every MCP client, sign-in with Google, one key per agent that you revoke from the dashboard.
Data
Servers in Germany. Never used to train models. Take any branch with you in a file, and delete everything whenever you want.
14 days free, then €19 a month.
A flat rate: no surprises on volume, and more agents don’t cost more.
Pro
- Unlimited memories
- All your agents, each with its own memory
- Claude, ChatGPT, Cursor and any MCP client
- Cancel anytime, from the dashboard
Card required; no charge before the trial ends.
Enterprise
- One memory for the whole company, by department
- Cross-department sharing, tracked and revocable
- Two-step verification required for everyone
- Direct support: you write, a person answers
Frequently asked.
Do I have to throw away my vector database?
No: fractal uses one inside and sits a layer above it. It recognizes when a new fact contradicts an old one and updates it, keeps the history of how information changed, and truly separates agents' memories.
Where does my data go?
To our servers in Germany, in the European Union. We never use it to train models and never sell it, and you can delete it from the dashboard whenever you want. From the graph you can export it, branch by branch, as a .fractal file. If you connect an outside model (OpenAI, Anthropic, Groq…), the text you send it goes to that provider: the dashboard tells you before the first message. For teams that want to keep it in-house, an app you install on your own computer or server is on its way.
Can an agent grant itself a permission?
No, and it's a ground rule: no tool ever changes a permission. If one did, a single malicious sentence in an agent's context would be enough to open another agent's memory.
What happens to a fact that turns out to be wrong?
It's replaced by the corrected version. The old one stays for a month as a safety copy, reachable by asking what was true on a given day, then goes away on its own. You can also delete it right away if you want.
What happens when the trial ends?
If you don’t cancel before, at the end of the 14 days the €19 a month subscription starts on the card you gave. Cancel from the dashboard, in Subscription, with one click.
Does it work with my framework?
If your agent speaks MCP, yes: Claude, Cursor, Codex, VS Code, Gemini CLI and the others connect with one command. For LangGraph there is a ready BaseStore, for Python a client, for PydanticAI an integration; without MCP there is the REST API.
Do more agents cost more?
No: with Pro all your agents, each with its own separate memory, fit in the €19 a month. Enterprise is for when the memory belongs to a company, split by department.
Two minutes, and your agent remembers.
14 days free, then €19 a month. Cancel from the dashboard whenever you like.