BCUB3

Nika OS hatch: persistent memory and a personal agent that specializes in its user's own work, inside a per-tenant frame, on Muse Glimmer 30B. General-purpose assistants forget, and they never get better at your job. At BCUB3, a small AI integration company in France, we run an operating layer (Nika OS) that gives each user persistent memory and agents that keep specializing on that user's real work. During this hackathon we rebuilt its core "hatching" loop from scratch as an open reference implementation on a Meta open-weight model. Everything in the repository was written during the event; the rest of Nika OS is context only. Built during the hackathon: - Per-tenant memory: one SQLite store per tenant, local embeddings, every memory carries its source and four trace ids. Isolation is enforced in code and tested (a store copied into another tenant is refused). - Orchestrator for headless Muse Glimmer 30B workers: memory retrieval recorded as a tool call, the tenant's checker verifies each answer, and every task is recorded as a trajectory (prompt, tools, output, verdict, ids). - Hatching: when a bottleneck is measured, a benchmark is pre-registered and hashed, a LoRA adapter is trained on the tenant's own trajectories, and it is promoted only if an exact sign test and a paired bootstrap (no normality assumption) both show a gain; otherwise it is rolled back. A negative control trained on shuffled answers goes through the same gate. Three levels of evolution: (1) prompt evolution in the spirit of GEPA (a reflector proposes which memory rule to pin into the prompt; implemented, not yet measured); (2) LoRA fine-tuning, measured below; (3) RL from the verifier's reward: today verifier-corrected SFT, policy-gradient RL is next. Real run (one rented GPU, bf16, 20.5 minutes, about 2.2 USD), 2 synthetic tenants, 240 trajectories and 100 held-out bench items each: negative control 0.57 -> 0.02, rolled back; machine-shop adapter 0.57 -> 1.00 (43 wins, 0 losses, p = 1.1e-13), promoted; consulting-firm adapter 0.05 -> 0.95 (90 wins, 0 losses, p = 8.1e-28), promoted. The specialized adapters also answer about ten times faster. A CPU replay mode reproduces every verdict from the recorded model outputs (that is the screen recording in the video). Honest limit: the tenants are synthetic, so the bench measures specialization to their conventions, not general capability. Related work by the author: two French patent applications filed with INPI (FR2500517, FR2513029), not implemented here. If you build on them, please cite them.