Episode memory
A small GRU tracks internal state across observations within an episode, rather than treating every decision in isolation.
Open research · v0.4
Learn. Simulate.
Verify.
A small neural brain for exploring learned state, action selection and bounded imagination. Train it on your CPU. Test it against reality. Inspect the evidence.
Research prototype, not a general reasoning system or a verified LLM enhancement.
01 / FlyBrain v0.3 architecture
A learned model, not another black box.
Explicit boundaries keep neural guesses separate from verified outcomes. The experimental FlyBrain v0.3 has recurrent state, learned action preferences, a budget head, and a latent one-step imagination module.
Bounded imagination
A trainable hidden-state predictor supports hypothetical rollouts. Simulation never substitutes for real execution.
Learned priorities
Policy and budget heads propose actions and 1, 2 or 4 thinking steps. For the v0.3 maze policy, budget labels are heuristic proxies. The separate v0.4 reward policy and budget head learn from verified toy-environment rewards, with a computation penalty for depth.
Observation
Recurrent memory
Action + budget
Imagined outcomes
Independent replay
Predictions are hypotheses. An external test environment supplies the ground truth.
02 / Evidence
Show the numbers. Keep the limits.
FlyBrain v0.3 results across three independently seeded trainings and 90 held-out toy mazes. These are real navigation completions, not model-declared wins.
3
training / evaluation seeds90
held-out toy mazesFlyBrain v0.3 · completed mazes out of 90
Adaptive depth did not consistently beat fixed depth across seeds (27 vs 23, 24 vs 28, 26 vs 23). The small aggregate difference does not establish a reliable benefit. A* remains the stronger baseline here.
New in v0.4 · Verified reward lab: A separate recurrent policy learned a tiny synthetic debugging workflow (240/240 held-out cases from three training seeds; untrained policy 0/240). A simple handwritten inspect → test → report rule also achieves 240/240, and forced one-step thinking also solves all cases. No advantage over the handwritten rule or fixed depth 1 has been demonstrated. This is a verified toy experiment, not real code debugging. No real source-code repair or coding sandbox was executed. The external adapter only proposes actions; it cannot execute them or grant rewards. Reward protocol ↗ · v0.4 evaluation report ↗
Maze results source: evaluation protocol and published reproducible reports. A toy maze benchmark is not a measure of general coding or LLM reasoning.
03 / Get started
Run it locally in minutes.
Python 3.10+ and PyTorch. CPU execution is supported; training time varies by hardware.
git clone https://github.com/Hookfabrik-AI/Latent-Brain.git cd Latent-Brain python -m venv .venv source .venv/bin/activate python -m pip install -e '.[test]' python -m pytest -q python -m latent_brain demo-fly \ --fly-checkpoint examples/pretrained_fly_v0.3.pt \ --checkpoint examples/pretrained_v0.1.pt
Run the v0.4 reward experiment ↗
Train
Train the v0.3 maze policy on generated environments and verified expert demonstrations.
FlyBrain guide ↗Use the local API
Request predictions or candidate plans from a localhost-only HTTP server.
API documentation ↗Reproduce results
Inspect the seeds, evaluation procedure, limitations and published reports.
Experiment protocol ↗Open research, honest claims. No general reasoning breakthrough is claimed. No network access or private agent integrations are required. Proposed actions are not proof of execution. Better results must survive held-out tasks and equal-budget comparisons.