Every session, your assistant wakes up with amnesia and re-reads your code from scratch. Cerebrofy gives the repository its own memory, and your AI asks it instead.
$ uv tool install cerebrofy $ cerebrofy init && cerebrofy build
Move where cross-session patterns get saved. Twenty minutes of work, you think. Your assistant opens a file. Then another. Then forty more, hunting for the answer the way you'd look for your keys by emptying every drawer in the house.
Watch the slide. That sweep is every file being read, in order, for one question. And tomorrow it will do the whole thing again.
Tokens per question on a 20,000-line codebase, if the assistant reads files.
Most of what it reads has nothing to do with the task, and it still pays attention to all of it.
Without the real call graph it guesses who calls what. Sometimes it guesses wrong, with total confidence.
The decision you made at lunch, the dead end it hit at three, the pattern it noticed. Gone by morning.
The model is fine. What it lacks is a memory.
A human brain never re-reads the world. It keeps a map, remembers what happened, works out why things are the way they are, holds goals, and has a feeling for how sure it is. Nature solved this long ago, with specialised regions that talk to each other.
So we borrowed the idea. Your repository gets one region for each kind of knowing, and your assistant asks the right region instead of searching everywhere. Scroll and watch each one light up.
Cerebrofy runs on your machine and keeps everything in one SQLite file. No server to host, no account, nothing uploaded. This is what happens when you run cerebrofy build.
Tree-sitter reads your code and names every function, class and module as a neuron. No LLM is involved, so the result is the same every time.
Calls and imports become typed edges: local calls, external calls and runtime boundaries.
Each neuron gets a meaning vector from a small embedding model that runs locally, stored with sqlite-vec.
Each module gets a written lobe summary, ready for an assistant to read in place of the code.
One MCP server per machine answers your assistant's questions, routed by whichever folder you are in.
Git hooks update the index on commit, push and merge, so you never have to remember to.
Ask "where do cross-session patterns get saved?" Cerebrofy finds the closest neurons by meaning, walks the call graph around them, and hands over just that: a few signatures, their neighbours and the lobe summaries.
On the slide, the bright points are the modules involved and the rest stay faint. Your assistant never opens them. The answer is about 10,000 tokens, with real names, real paths and real call chains.
Same task. Your assistant asks one question and gets the three neurons involved, who calls them, what was decided about them last month, who owns that lobe, and how sure the index is. It starts with the answer, and so do you.
Two commands. It builds locally, hooks itself into git, and your assistant can start asking.
$ uv tool install cerebrofy $ cerebrofy init && cerebrofy build