Skip to content

2.5.0 — An audit of one real brain, and what it showed

Choose a tag to compare

@Octonove Octonove released this 20 Sep 01:55
· 120 commits to master since this release

An audit of one real brain — 1,200 neurons, 97 sessions, five months of daily use — and of the machine it runs on. Nothing here came from a whiteboard: every item started as a number that looked wrong. Full detail in the changelog.

The brain backs itself up. npx crbro-memory backup | list | restore, and crbro_consolidate makes one a day on its own, rotated, beside the brain it belongs to. The quarantine and machine tokens never travel. CRBRO_BACKUP_DIR points it at a synced folder.

Dates that mean something. Entries written before 1.13 had no date (249 of 457 on the reference brain). crbro_maintenance backfill_dates recovers them from the date stated in their own text — never guessed: the neuron's created was measured as a fallback (a median window of 117 days) and thrown away. Recall now breaks ties towards the newer telling (a lift of at most 4%; on 115 real queries it only ever flipped exact ties), and crbro_recall since / kind narrow a search to "the last two weeks" or "only past mistakes".

Housekeeping that reports before it acts. Every maintenance run lists entries whose own deadline has passed, neurons that outgrew one read, and the one-line neurons a bulk import left behind. crbro_revise move_to splits a neuron keeping every date; compact:true folds an import's leftovers into a digest (754 neurons → 1 in 6.8 s on the reference brain, real queries unchanged).

Memory at the moment of action (opt-in). npx crbro-memory install-hooks --guard: before a shell command runs, the stored errors, debts and patterns that mention that command are added to the model's context — three at most, once per session, never blocking. Recall only answers when somebody asks; nobody asks one second before firebase deploy.

One process owns the brain (opt-in). npx crbro-memory daemon on: the clients of a brain stop loading a copy of it each and become 55 MB proxies to one daemon. Three clients: 2,153 MB → 1,140 MB, the second client ready in 0.3 s instead of 1.8, and a line saved in one chat recalled in another at once. It can only ever cost speed: no daemon, or a daemon that dies mid-call, and the client serves itself and carries on. Reviewed adversarially before shipping — fourteen findings, each now a test.

Two experiments, pre-registered. Synapses in the ranking: measured on 633 queries, no effect, discarded — the code was removed rather than left asleep. The agentic benchmark LIMITS.md designed is now built (frozen tasks, tested scorer, isolation canary) and not yet run; no figure from it exists anywhere.

419 tests (312 before this release). The blind retrieval benchmark is unchanged: 71% / 77% / MRR 0.744 for the keyword engine.