Blood glucose, insulin, sleep, heart rate, and context arrive as one messy moment.
A Type 1 diabetes example moves through Chamelia. Observations become a perceived state; memory and planning shape an action; the observed outcome returns to the system as learning.
Read the episode as text
Blood glucose, insulin, sleep, heart rate, and context arrive as one moment.
Perception compresses those observations into one compact state.
That state queries memory, which retrieves a similar morning, poor sleep, and exercise sensitivity.
The compact state and those labeled memories enter the configurator from either side.
The actor draws three named candidates from its bag of actions. Each becomes the root of its own imagined future; cost scores the results while metacognition allocates attention and compute.
The lowest-cost valid candidate leaves search and becomes the chosen action.
Reality returns an observed outcome, which may differ from the prediction.
The outcome sends an upgrade pulse through the system; sleep consolidates what worked and returns a reusable skill to memory.
Chamelia began as a Type 1 diabetes project.
Someone close to me has Type 1 diabetes.
The same number could mean something different depending on sleep, exercise, stress, hormones, illness, or what had happened over the previous few days. We started prototyping a system that could keep track of those observations and use them to recommend changes to insulin-pump settings.
That project eventually became Chamelia: a system designed to remember what happened, understand the current situation, consider what different actions might lead to, and learn from the outcome.
The working research stack
What is actually inside the loop.
Perception / JEPA
Perception is built around H-JEPA, drawing on Yann LeCun’s joint-embedding predictive work. It learns compact representations from context at several scales without wasting effort reconstructing every raw observation. The VICReg objective helps keep those representations varied instead of collapsing into the same answer.
World models / RL
The planning side draws from model-based reinforcement learning, especially the ideas behind Dreamer and MuZero: learn a useful state, predict what actions may do to it, and reason inside the model before acting. Chamelia pairs an action-conditioned world model with tree search rather than treating the first plausible action as the answer.
Unified memory
Episodic memory keeps whole moments, including the options considered and what happened afterward. Semantic memory distills beliefs across moments. Procedural memory stores methods the actor can call again. Testimony and injected knowledge can enter memory too, but remain evidence to evaluate rather than instructions to obey.
Sleep / skill formation
LOVE finds patterns that kept working. Stitch compresses a successful sequence into a reusable move. RSD turns missed opportunities and regret into candidates for next time. BodeGEN explores variations so the system can keep the version that tests best.
Sleep / promotion
Choreographer replays past situations to ask whether a candidate skill would actually have helped. LILO gives useful skills readable names. Audit promotes what earns its place and drops what does not. Consolidate turns repeated evidence into beliefs.
Constraints and cost
Constraints remove actions that are not allowed before search. Cost scores what remains with both readable domain rules and learned long-run judgment. What is possible, what is permitted, what might happen, and what matters stay separate enough to inspect.
For builders, researchers, and aligned investors
Bring a hard problem. Help test the thesis.
Iām looking for technical collaborators who want to pull the architecture apart with care, and investors who understand that long-horizon intelligence is an engineering problem worth staying with.
