How memory works

Fibonna keeps one memory per project, on a stack that is yours alone, that every agent you connect writes to and reads from. Your agents save notes, decisions and findings as they work. Fibonna compiles those into knowledge pages and lessons. The next session, in Claude Code, ChatGPT or any other tool you have connected, recalls them with the source attached, or says "not found" instead of guessing.

What gets remembered

Four kinds of things, all written by your agents through the connector:

Notes, decisions and findings. "Remember that we chose Postgres over MySQL because of its JSON support." Each note is saved in full and tagged with the project, sprint and task you are on.
Handoffs. At the end of a session, "hand off" writes down what changed, what you decided and what comes next. The next session starts from it.
Lessons. After a handoff, Fibonna extracts what the session taught you, checks it against the lessons it already holds, and keeps the distilled version. A mistake is written down once, so your agent can ask "have we hit this before?" before it repeats it.
The work itself. Projects, sprints and tasks, with their status, so "what is left?" has an answer that does not depend on anyone's memory.

Fibonna does not read your repository or watch your terminal. What it holds is what your agents write to it. The privacy page lists, in full, what leaves your stack and who receives it.

How it is compiled

Notes pile up. Left raw, they are a search problem: a hundred saved decisions about one feature, some superseded, some contradicting each other. So Fibonna compiles. It groups the notes on one topic into a knowledge page, rewrites the page when a new note changes it, and updates the related pages that the change touches. Your agent can list the pages that exist before it reads one, so it sees the shape of what the project knows.

Compilation runs on your stack, calling the AI provider you chose with your own API key. The text sent is the note being compiled and the pages it may update, nothing else. Embeddings, which make search by meaning possible, are computed on your stack and never leave it.

What comes out is three things an agent can act on: decisions (what was chosen and why), lessons (what to do differently next time) and open loops (what was started and not finished).

How recall works

Three ways, all through the same connector:

At the start of a session. Your agent opens a session and Fibonna loads what it remembers about the project: the last handoff, the open tasks, the pages that matter. You do not re-explain yesterday.
On demand. "What did we decide about the pricing page?" searches everything saved to the project by meaning, not only by exact words, and returns the matching notes and pages. For a task in hand, one call gathers the task, its sprint, the related notes and the lessons that apply.
Before a mistake. "Have we hit this migration problem before?" searches the lessons.

Every answer points to the note or page it came from, so you can open the source and check. When nothing matches, Fibonna says "not found". It does not fill the gap with a plausible guess, because a confident wrong answer costs more than an honest empty one.

Across tools and machines

The memory lives on your stack, not in any one tool and not on any one machine. Connect Claude Code on your laptop and your desktop, add ChatGPT for the conversations that happen away from the code, and they read and write the same memory. A decision made in a chat on Monday is there when your coding agent starts on Tuesday, on whichever machine you sit down at.

Keep the instruction files your agent already loads at startup; they are the right place for rules you write by hand. Fibonna holds what outgrows them: the decisions, lessons and open loops that accumulate over weeks of work.

Export and leaving

Everything is yours. You can ask to see, export or delete everything Fibonna holds for you, and we do it within 7 days. When you leave, we delete your whole stack, and backups of a deleted stack are purged within 30 days. The privacy page has the details; the request goes to [email protected].