ocbrain

Local · source-backed · MCP

Memory your agents can inspect.

OCBrain gives Codex, Claude Code, Hermes, OpenClaw, and compatible MCP clients one local evidence and task-continuity layer—without turning chat history into unquestioned truth.

Fresh contexts reorient from a deterministic briefing, retrieve only what matters, inspect exact sources, and leave verifier-backed outcomes for the next agent.

Verified 2026-09-01 · Current main is green at 01e94b2.

1,133tests passed locally on current main
32 KBhard context-packet ceiling
13tools in the normal MCP profile

The useful loop

Retrieve less. Verify more. Learn from outcomes.

OCBrain is not a second agent and not a bag of embeddings. It is the evidence layer beneath the agents you already use.

01

Resume deterministically

brain.briefing returns open goals, done and failed attempts, the latest closeout chain, and standing gotchas in at most 1,500 characters.

02

Retrieve, then inspect

brain.context stays bounded. When a decision needs more, an issued handle expands exact material with scope and content-hash checks.

03

Leave a usable trail

Feedback records whether retrieval mattered. Closeouts connect context to actions, verifiers, outcomes, and the next task iteration.

Start small

One checkout. One client. One real round trip.

A fresh brain starts empty. Install current main, initialize an owner-only database, connect one compatible client, and verify actual tool use before adding anything else.

The latest packaged release remains v1.1.0; the harness, supersession, and thirteen-tool contract are newer unreleased main capabilities. OpenClaw and Ollama are optional.

Open the complete install and upgrade guide
git clone https://github.com/jonathangu/ocbrain.git
cd ocbrain
python3 -m venv .venv
.venv/bin/pip install -e .
.venv/bin/ocbrain init
.venv/bin/ocbrain doctor

Current main · verified 2026-09-01

Small context, durable progress.

Current main adds a deterministic harness surface on top of source-backed retrieval: agents can resume without replaying transcripts, check whether work already succeeded or failed, and keep executable finish lines in versioned repository specs.

Deterministic startup

brain.briefing is scope-bound, byte-stable for the same corpus state, and hard-bounded to keep reorientation small.

No duplicate implementation

brain.ledger groups verified, failed, and in-flight attempts by stable task reference, independent of search phrasing.

Procedures earn promotion

Optional local DAG mining can mint high-support gotchas. Weak procedure candidates abstain instead of becoming instructions.

Deliberate boundaries

Useful memory should not acquire ambient authority.

  • Raw transcripts remain evidence, not automatically served truth.
  • The core installs no scheduler, timer, pager, or recurring maintenance.
  • Hosted delivery is filtered before ranking; local paths stay local.
  • Corrections and tombstones are later events, never silent rewrites.
  • Bundles are explicit operator file operations, not agent tools.

Go deeper

Use the page that matches the job.