Independent research project
Neum is a long-lived AI agent that combines neural and symbolic cognition: meaning-based persistent memory, a compressed symbolic language, calibrated uncertainty, multi-strategy reasoning, an imagination engine, self-improvement, and full-duplex voice — engineered and evaluated as a research program, not a demo.
One codebase, running continuously — cognition built from a stack of specialised modules, with a health registry watching every one of them and every capability independently switchable.
A Neuro Semantic Pattern Code keeps memory the way a person holds a concept, not a transcript: one meaning can span many words, so a whole idea compresses into a handful of symbols. Facts, frames, and relations survive across sessions — the agent wakes up tomorrow still knowing what it learned today.
NSPC Space runs in the shadow of the primary store. Once it has proven itself with enough data, it ascends and replaces the primary store — which is archived, never erased.
299 seed neums are only the beginning: the dictionary evolves. EvoRules let knowledge carry confidence that rises and decays with use and success, and composites used often enough compact into brand-new symbols during sleep. Meaning drift is tracked; polysemy is verified before it is trusted.
Uncertainty is computed from observable signals, not guessed. Asked about something it was never told — "what is my mother's name?" — the agent abstains and says so, instead of confabulating. Calibration is measured (ECE 0.052), not assumed.
Mental simulation across physics, social, causal, temporal, and counterfactual scenes, at selectable fidelity — scenes the agent can hold in mind, not just strings it can repeat.
A coordinator chooses between dual, abductive, creative, and recursive strategies — and learns the choice from feedback. Answers are cross-checked against stored relations, so a contradiction lowers confidence instead of slipping through.
One episode is stored per interaction. Insights are surfaced at most once a day, then judged. Perceived valence becomes emotional context, and the user's visible state enters the conversation.
The agent trains itself on a measured curriculum during sleep cycles, and every step is recorded in an improvement trajectory. Progress is measured, never asserted.
Where attention goes is itself a decision: stimuli become evidence, hypotheses, and competing goals, then a stateful attention plan allocates processing power. When idle, intrinsic drive keeps the agent thinking instead of waiting.
Perception starts at an event-based retina — on/off cells, edges, motion, like a biological eye — with foveated attention instead of flat pixel processing. The voice loop is full duplex: it hears while it speaks, and can be interrupted mid-sentence.
Linear, exponential, periodic, trend, and anomaly detection across data, backed by statistical, sequential, and structural feature extractors.
The interesting question is not whether an agent can do something once — it's whether a change actually helps. So the project runs on experimental discipline:
The full evaluation harness runs end-to-end without a paid model, so most development happens at zero cost. The agent's abstention suite correctly refuses 8 of 12 unanswerable questions while keeping calibration tight — behavior that appears only when memory and uncertainty are genuinely wired together.
Some work can only be done with a real LLM behind it: semantic-mutation thresholds, polysemy verification, recursive-reflection quality, and the live full-duplex voice loop. Those sessions currently run on rented internet-café time — the one bottleneck between this project and steady iteration.
Neum is built by one person in Indonesia — no funding, no team, no institution. Claude has been the primary tool throughout the project: Claude Pro carried the build until the budget ran out, and the roadmap since then has been paced by compute, not by ideas.