The earlier fork integration gave TanyaOS a maintained BrainCog dependency. The next architectural decision changes that direction: build and own the active cognitive implementation.

TanyaOS now runs original source in BACKEND/brain/cognition. Neither the original BrainCog checkout nor the maintained fork is required to start, test, or package the application. The fork remains a separate historical research project with its original licenses and credits.

Learning connected to real cognitive events

The new NumPy engine implements sparse leaky integrate-and-fire dynamics, deterministic event encoding, eligibility traces, and bounded outcome-driven weight updates. Its configuration assigns exactly 10,000 units to twelve ensembles: sensory input, attention, workspace, episodic association, semantic association, self, affect, goals, prediction, action, skills, and reflection.

Inputs, decisions, and observed outcomes advance the network. Measured activity also follows local model inference and speech callbacks. This is software computation with inspectable state, not a biological measurement or evidence of subjective experience.

Outcomes update empirical success predictions and neural weights. Candidate ranking considers those predictions, risk, resource cost, and learned neural associations. The protected cognitive kernel still owns identity, consent, policy, and executive authority. A learned preference or successful previous action cannot grant a missing permission.

Continuity belongs to the local system

Eight memory classes share the existing account-scoped cognitive database. Source IDs, confidence, approval, suppression, and links remain available for inspection. Selective recall uses a token index and bounded candidate set, allowing an older relevant episode to be retrieved without injecting the entire archive into a language-model prompt.

Neural checkpoints preserve learned weights and traces, validate configuration and checksums, and retain a known-good previous state. The latest checkpoint payloads also reside in SQLite, so the existing database backup can restore neural learning alongside cognitive records.

Repeated explicit communication preferences can gradually adjust response detail within small bounds. Recorded revisions are reversible and cannot rewrite core identity. Approved procedures reuse registered skill actions; execution checks every step's permission and schema before the first side effect, then records the actual outcome.

Evidence and limits

The combined backend regression run passed 90 tests across original cognition, canonical processing, offline operation, accounts, recovery, agentic execution, native desktop skills, and monitor HTTP behavior. Focused learning tests demonstrate weight changes that survive restart, improved predictions after repeated outcomes, changed later candidate selection, account-isolated recall, checkpoint recovery, and blocked execution without permission.

Renderer lint, the production build, action-contract checks, and all 16 Electron runtime tests also passed. A small CPU benchmark measured roughly 24–51 milliseconds per sixteen-tick sample depending on concurrent load on this machine. Network arrays occupy 1,270,000 bytes; this excludes models, Python, storage, and graphics.

This is a working first implementation of the larger architecture. GPU execution is not implemented. The world model currently records empirical task/action hypotheses, personality learning covers response detail, and consolidation is operator-requested bounded replay. General causal reasoning, automatic idle scheduling, broader trait learning, long-duration evaluation, visible Windows/audio QA, and signed update installation remain open.

The source implementation record documents those boundaries and the remaining gates. Retained third-party atlases, speech components, and other assets continue to carry their licenses and attribution. An original cognitive engine changes ownership of that implementation; it does not erase the provenance of other components.

Explore the TanyaOS project and its source repository.