Local rules. Live experiments.
Inspect the mechanisms behind cellular memory and causal language. Running locally, powered by ZIPP.
Cellular memory
Select a compute backend and start the engine.
No server, account or model API required.
Ready when you are. Nothing is running yet.
The answer has to travel.
The original model learns key and value representations; a supplied local delta rule writes private matrices. A fresh query moves one cell per update and reads only those matrices. These are real checkpoint computations, not a recording.
Local write law and routing are supplied. Changing private memory is not training the shared weights.
Read the research context ↗Reproduce & inspect the source ↗Observed associations
Reference labels for this episode. Query code never receives these values.
Start the engine to create the first episode.
Inspect the Python that runs in ZIPP
Actual bundled source. Matrix graphs use the selected backend; this viewer is read-only.
Download memory.pyStart the engine to view the verified source, or use the download link above.Explicit backend. No silent substitution.
WebGL2 needs supported hardware and floating-point render targets. CPU reference is an intentional fallback, not a GPU result.
Live state, not a benchmark claim.
Colours visualise actual state. Per-command timings include host overhead and readback; they are not hardware comparisons.
Original checkpoints. Bounded sessions.
Python runs in a worker with execution limits. Stop engine terminates the worker and discards the current state.