How sharm gets smarter
A large language model reads the world. sharm learns to live in it — remembering, feeling and anticipating, and sharpening on your world with use. Explore what makes it a different kind of intelligence, and the measured signal behind it.
Not a better chatbot — a different shape of mind
LLMs like Claude and GPT spike on language. sharm is built for the experiential axes an LLM structurally lacks — memory, affect, anticipation, privacy. Hover an axis to see the difference.
Memory
Forgets once the context window scrolls past — recall is retrieval bolted onto frozen weights.
A hippocampal loop stores and replays episodes, so yesterday shapes today.
Illustrative capability profile — what each system is designed for, not a head-to-head benchmark score.
The more of your world it sees, the sharper it gets
Every moment runs one loop — perceive, remember, feel, anticipate — and each pass leaves the model more tuned to your life. A general model is frozen after training. sharm keeps learning.
Perceive
See and hear the moment, on-device.
Remember
Bind it to episodic memory.
Feel
Weigh how much it matters.
Anticipate
Project what comes next.
Proof it learns: episodic memory, sharpening
Measured · recall@1As the memory module trains on grounded data, its ability to recall the right past moment climbs from near-baseline to near- ceiling — on our leak-guarded probe (chance ≈ 1.6%).
Numbers we can stand behind
Results from our own evaluations — leak-aware, on held-out data. We report the capabilities that matter for a world model: reading affect, recalling the past, recognizing people and products, and learning efficiently.
Affect correlation (Pearson r)
Measuredvs distilled human judgment · higher is better
Training loss · one $0.52 run
Measured · −37%10-head joint trainer · H100 · lower is better
Results are from sharm's own evaluations on held-out data. sharm is a world model — a different category from general chat LLMs — so we benchmark the capabilities that matter for living in the moment, not general language tasks. Edge-device timings are simulated until on-silicon evaluation is published.