Your files, your ground,
an AI that remembers.

The problem

Your agent does not fail loudly. It forgets. You re-explain the same constraints every session — your conventions, the approach you already rejected and why, what “done” means here. That knowledge lives only in your head, and you pay for it again every day.

A correction that is only applied is spent; a correction that is written down becomes capital. That is the whole difference between an agent that obeys and an agent that learns.

The loop

1. Give it your filesYour notes, your projects, your decisions. Where they already live.
2. EncodeOne fact per file, with a description that doubles as the index hook and as embedded keywords.
3. RecallSemantic search surfaces the right notes at the start of the task — in any language.
4. CorrectEvery correction becomes a permanent rule, carrying its reason. This is where memory starts to compound.

The engine, open source

A folder of markdown notes, indexed continuously, served over MCP — plus a selfcheck that verifies the memory is still there. Because it fails quietly: the session boots normally and simply knows less, and you put that down to the model.

git clone https://github.com/ninabot-ch/corthexis
python -m corthexis.index
claude mcp add -s user corthexis …

The repository on GitHub →  Apache-2.0, no service to sign up for.

The corpus the method came from

Produced while actually operating a company, not by filling a demo folder. This corpus is private and will not be published: it is full of infrastructure and customers. The engine is entirely open — and the proof that matters is not our memory, it is what happens when you run the engine against yours.

The cohort

Six live sessions, from Geneva and remote, spent encoding your files — not sample data. Small group, first session in October 2026.

No seat is sold from this page. You get the dates and the price once they are set.