Resume small. Retrieve narrowly. Verify the finish.

RAG is not dead here—and long context is not the database.

OCBrain's history is effectively unbounded and keeps changing. Re-reading it all for every question would waste compute and dilute attention. Pure top-k snippets are not enough either: models need a coherent local picture to compare sources, detect omissions, and understand what changed.

The index is replaceable infrastructure. The evidence and outcome ledger is the product.

The fresh-context loop

1. Briefing

   brain.briefing deterministically returns open goals, done and failed attempts, the latest closeout chain, and gotchas within 1,500 characters.

2. Ledger

   brain.ledger checks a stable task reference before search can convince an agent to build something twice.

3. Context

   brain.context returns a focused dossier and retrieval-use ID. Goals still point to versioned repo specs and executable finish lines.

4. Source

   brain.source expands one issued handle in the same scope and verifies that the body has not changed.

5. Work

   The runtime treats context as evidence, not orders, verifies conflicts, and acts only within its existing authority.

6. Feedback

   brain.feedback records whether the retrieval was helpful, used, irrelevant, ignored, or harmful.

7. Closeout

   brain.closeout appends status, decision impact, retrieval links, artifacts, verifiers, actions, outcomes, and provenance.

The default runtime profile exposes exactly thirteen tools: briefing, ledger, goal open/close, context, source, search, digest, get, feedback, ingest, closeout, and supersede. The explicit admin profile adds six local lifecycle/review controls for nineteen total. It adds no hosted judgment, hosted embedding, training, scheduler, watchdog, or bundle-transfer tool.

Dense retrieval is optional, local, and disposable.

v1.1.0 can combine lexical and dense candidate ranks with reciprocal-rank fusion. The dense index lives in a separate owner-only SQLite sidecar and uses only a loopback Ollama endpoint. It records the core event head, model digest, dimensions, and query-instruction hash so incompatible or stale vectors do not quietly participate.

Lifecycle, scope, visibility, and delivery checks create the eligible inventory before either lexical or dense candidates are ranked. Retrieval quality features cannot widen authority.

One semantic authority

brain_events is the append-only hash-chained authority. A correction, scope promotion, or tombstone is a later event; history is never silently rewritten.

The strict v1 core contains no legacy relational knowledge/evidence plane and no dataset, trainer, judge, embedding, autopilot, scheduler, watchdog, deadman, or pager tables.

Scope controls affinity and privacy.

Each object carries scope type, scope ID, visibility, egress policy, and provenance. Local retrieval ranks every non-secret eligible scope by affinity and reports the actual scope_mix; the deprecated cross_scope flag no longer widens anything. Confidentiality remains a hard gate. For the ordinary hosted_model target, only non-confidential hosted_ok objects in the explicit scope list are eligible.

The server chooses the delivery target; tool arguments cannot downgrade it to local_model. Local packets report the eligible inventory and actual scope mixture. Hosted delivery stays explicitly scoped and reports delivery exclusions without exposing any excluded ID, body, path, or object sample. Every complete serialized packet is hard-bounded to 32 KB.

Direct get and source expansion enforce lifecycle, quarantine, confidentiality, and delivery boundaries, so knowing an object ID is not an access bypass. Hosted source handles use ocbrain:// locators instead of local paths and expand only in their original scope.

Migration maps only explicit legacy context. The placeholder value workspace becomes conservatively quarantined legacy_unscoped; it is not promoted into a fictional global project.

Historical at_ts retrieval is explicitly rejected in v1 until it can be implemented against event sequence correctly. It is never silently approximated.

Cross-machine exchange is evidence transfer, not brain cloning.

An unbounded private lake still needs backup and migration choices, but automatic replication would widen the privacy and authority surface. v1.0.1 therefore makes portable exchange an explicit file-only operation.

A click is not a universal reward.

The same physical action can mean different things on different sites. OCBrain keeps both a portable shape and the local meaning.

An experiment may derive one scalar reward for one policy. The append-only closeout keeps the components so a later model can reinterpret them without assuming a click, signup, subscription, deploy, or test pass means the same thing everywhere.

Procedural memory is earned from traces.

OCBrain's optional local procmine path normalizes tool-call traces into privacy-scrubbed episodes and mines repeated DAG shapes. A procedure is not promoted because one run looked plausible: the miner requires at least five episodes, a shared subsequence of at least three steps, and at least 50% task-family coverage. Below the floor it records an abstention.

One distribution means one auditable boundary.

ocbrain is the whole product: ledger, projections, retrieval, MCP, receipts, backup/migration, curation scripts, diagnostics, and the public-safety scanner. The former ocbrain-training and ocbrain-ops companion distributions were deleted after their stores remained empty. There is no external CLI extension mechanism and no separate package installation.

Migration preserves before it transforms.

core-migrate-v1 creates a coherent immutable archive, preserves the exact event sequence and hashes, appends deterministic import events, extracts training and operations stores, bulk-rebuilds FTS, verifies counts and integrity, and publishes only fresh outputs plus a manifest.

The July 13 full rehearsal verified 671,580 events, 232,677 evidence objects, 137,530 beliefs, 241,185 evidence links, 108,222 serving/search documents, 1,725 retrievals, and four closeouts. It made no hosted or network call, started no scheduler, and did not activate or repoint the live database. The operator later activated a byte-identical live copy provisionally; fresh Codex, Claude Code, and OpenClaw processes all passed context → source → feedback → closeout, so the pointer was retained.

Training is not a runtime capability.

The current distribution contains no trainer, dataset pipeline, hosted judge, or prepared training pack. Adding that authority would require a reviewed source change; no credential or configuration flag can silently enable it.