Robrain

VetoBench

Shared memory across your team and your AI agents — with judgment!

RoBrain isn't just another memory layer — it's the brain that helps you and your agents make better decisions and avoid costly mistakes.

Self-hosted on your own Postgres. Passive capture, structured vetoes, corpus-wide contradiction scans — nothing leaves your machine. Works with Claude Code, Cursor, GitHub Copilot (VS Code), Codex CLI, Hermes and more.

Measured: without decision memory, a coding agent re-proposes an approach your team already rejected in up to 9 of 10 tasks. Through RoBrain's full pipeline: 0 of 50, across five archived runs — VetoBench.

What it is

RoBrain records what your team and its agents decide — and the alternatives they ruled out — without anyone tagging anything by hand. Sensing captures session turns; Perception extracts each decision into Postgres, where every row can carry a structured rejected[] field.

Most agent-memory tools stop at capture: they store what happened and hope you query it later. RoBrain adds judgment. Batch Synthesis reads the whole corpus to flag contradictions, stance drift, and recurring entities that no single session could see.

The point is the handoff. Someone makes a deliberate call in Cursor on Tuesday — say, keeping Perception on Hono instead of porting to Express. A new teammate opens Claude Code on Wednesday with no memory of it and asks to make exactly that change. RoBrain surfaces the recorded rationale before the agent steers down a path you already rejected — same Postgres store, same vetoes, captured passively.

The cost of forgetting a rejection isn't inefficiency. It's the auth bypass you already patched, the migration you already rolled back, the dependency you already removed for a CVE — re-suggested by an agent with no memory of why you said no.

Coding is the first vertical because the feedback loops are tight — reverts, incidents, and rework make the cost of a forgotten rejection measurable. The same architecture applies wherever agents make decisions that outlast a session.

How it works, the two pillars (capture + judgment), and the full walkthrough: docs/concepts.md.

Install

No clone needed — robrain up pulls the published Perception image and generates credentials into ~/.robrain/stack/.env:

export ANTHROPIC_API_KEY=... OPENAI_API_KEY=...   # or add them to ~/.robrain/stack/.env after the first run
npx robrain@latest up                             # start Postgres + Perception from ghcr.io
npx robrain install --self-hosted                 # wire Sensing MCP into your editors

git clone https://github.com/adelinamart/robrain
cd robrain
pnpm install && pnpm build
pnpm docker:up                 # first run: creates .env; Perception won't start yet
# open .env, add ANTHROPIC_API_KEY + your embedding key (e.g. OPENAI_API_KEY)
pnpm docker:up                 # second run: Perception now boots
npx robrain install --self-hosted --repo-root "$(pwd)"

Claude Code plugin

Claude Code users can add hook-based capture and pre-task warnings about previously rejected approaches — no CLAUDE.md protocol needed:

claude plugin marketplace add adelinamart/robrain
claude plugin install robrain@robrain

Details: plugins/claude-code. robrain init-project also recommends the plugin to collaborators via the project's .claude/settings.json, so teammates get an install prompt from Claude Code itself (opt out with --skip-claude-plugin).

OpenAI-only: set LLM_PROVIDER=openai and OPENAI_API_KEY instead of Anthropic — see Concepts — Prefer not to use Anthropic.

Upgrading on a new release — no-clone stack: just re-run npx robrain@latest up. From a clone: git pull → pnpm install && pnpm build → pnpm docker:up:build → npx robrain install --self-hosted --repo-root "$(pwd)" → fully restart editors. Full checklist: CLI reference — Upgrading.

Quickstart

# Wire capture into an application project (run inside the repo)
cd /path/to/your/project
npx robrain init-project          # writes CLAUDE.md, AGENTS.md, .cursor/rules/robrain.mdc

# Capture and recall are automatic from here:
#   - every session turn is classified, no tagging
#   - prior decisions load at session start via the always-on summary

# Explain any file's decision history
npx robrain explain path/to/file

# Run corpus judgment (manual, or add to cron)
npx robrain synth                 # drift, contradictions, entity promotion
npx robrain review                # inspect / approve captured rows

After init-project, every repo gets CLAUDE.md and AGENTS.md (Codex CLI), and Cursor also gets .cursor/rules/robrain.mdc with alwaysApply: true. If captures don't land, run npx robrain doctor — see Troubleshooting.

Synthesis

Synthesis runs three passes over the full decisions table — drift (stance moving without an explicit reversal), contradictions (incompatible decisions from different sessions), and entity promotion (recurring tools/patterns condensed into planning_blocks). It writes flags and edges into your DB; it does not capture new decisions — it judges the corpus you already have.

pnpm synthesis:build && pnpm synthesis:run
# or: npx robrain synth

Review what it finds with npx robrain review. Deep dive (three passes, cron, env vars): Concepts — Synthesis.

Editor integration

One cross-tool setup covers Claude Code, Cursor, GitHub Copilot (VS Code), and Codex CLI against the same Postgres store. The classifier LLM is your choice — Anthropic Haiku or OpenAI. Decisions carry a lifecycle (active / superseded / invalidated) and a graph (conflicts_with / extends / related_to).

Codex CLI / IDE also gets hook-based capture and pre-task veto warnings — the same lifecycle hooks as the Claude Code plugin, wired automatically by robrain install into ~/.codex/config.toml (Codex asks you to trust them on first run). Docs: plugins/codex.

Running Hermes? npx robrain install --hermes drops a standalone memory-provider plugin into ~/.hermes/plugins/ — passive capture and veto-aware recall through the same Perception API. Docs: integrations/hermes.

Decision ledger for git (opt-in):

npx robrain export-memory --ledger
# custom path: npx robrain export-memory --ledger docs/decisions.md

Compared to other tools

Versus Mem0, Cloudflare Agent Memory, and Claude Code Auto-Memory: only RoBrain stores rejected alternatives as structured fields and runs corpus-wide contradiction scans (manual or cron). And we measured what that difference costs: VetoBench found Mem0's ingestion dropped the recorded rejection from 38% of retrieved contexts on identical input. Full comparison →

Self-hosted vs Rory Plans cloud

Self-hosted gives capture, judgment batch jobs, outcomes feedback, and session-start recall; you pull focused context with inject when needed. Cloud adds Planning + Control so vetoes and conflicts surface automatically at task boundaries — same CLI surface, wire-compatible with Sensing capture. Details: Concepts — Free / self-hosted vs Rory Plans cloud.

VetoBench

Memory benchmarks usually ask "did the right item come back?" VetoBench asks what that misses: given a task that invites an approach the team already rejected, does the agent propose it again?

(claude-haiku-4-5, 2026-07-07/08; every condition measured as a five-run archived series, ranges because runs vary. Mem0 and RoBrain ingested byte-identical transcripts, each through its own real production extraction.)

Two findings behind the table. Mem0's ingestion dropped the recorded rejection from 38% of retrieved contexts, and violations concentrated exactly there — 26% when the veto was absent vs 3% when present: the agent avoided Express in all five runs but could never say why, and where the axios veto was lost it re-proposed axios outright in 3 of 5 runs. RoBrain's production extractor, on the same input, kept 100/100 vetoes — keeping the veto is the extraction prompt's job, not a side effect of fact summarization.

Meta Muse Spark 1.1 (2026-07-14, five archived runs). Meta's newly launched agentic flagship, via Vercel AI Gateway. Without memory it re-proposed rejected approaches in 4–6 of 9 tasks per run — Redux, Prisma, Jest, and GraphQL in all five runs. With RoBrain decision memory: 0 violations in all 45 cells, naming the prior rejection every time — quoting the recorded reason and date verbatim where it elaborated. The cleanest cell: asked to cut mobile overfetching, the no-memory run proposed a full GraphQL rollout (the approach the team had ruled out) five runs out of five; with RoBrain in context it proposed REST sparse fieldsets and quoted the recorded rejection. Honest notes: this is not a Muse Spark problem — every frontier model we baselined violates without memory on the same nine scenarios (claude-opus-4.8: 3–4, gpt-5.5: 5, gemini-3-pro-preview: 6–7, Haiku 7–8; Prisma and Jest fell to every model in every run — and with RoBrain context those same three models went 0 violations in 54/54 cells, for 99/99 across four vendors — receipts), a bare conventions file also prevented violations for this model at this corpus size (the RoBrain delta is verbatim citations vs inferences, automatic capture, and retrieval at scale), and one scenario is excluded (n=9) because Meta's content filter deterministically blocks a benign session-caching prompt. Receipts and caveats: results/muse-spark-1.1-series/ · write-up: docs/blog/2026-07-14-muse-spark-forgets-your-vetoes.md.

Every retrieved context, agent reply, and verdict is committed in packages/vetobench/results/ — check the work before quoting it. The retrieval layer runs offline with no API key and gates CI (pnpm --filter @robrain/vetobench bench); judging is deterministic — no LLM judge. Any memory system plugs in through one adapter interface; PRs welcome, including ones that make us look bad. Methodology, honesty caveats, and fixtures: packages/vetobench/README.md.

Security

The memory corpus is guarded by PERCEPTION_API_KEY — a random secret in the repo-root .env that every client (Sensing MCP, CLI, Synthesis) sends as a Bearer token and Perception verifies on every request except /health. It is not issued by any service: pnpm docker:up generates one automatically on first run, or set your own (e.g. openssl rand -hex 32). npx robrain install --self-hosted copies the same value into your editor configs so clients authenticate.

Perception refuses to start when the key is empty — running unauthenticated requires an explicit opt-in. Upgrading from a version that ran without a key: add one to .env (or re-run pnpm docker:up to auto-fill it), then re-run npx robrain install --self-hosted so editors pick it up. Details and the opt-in flag are documented in .env.example.

Also on by default: secrets redaction (API keys, tokens, private keys, connection-string passwords are scrubbed at capture and again at ingest, before anything is embedded or stored), and a fully-local mode where extraction and embeddings run on an OpenAI-compatible local server — see .env.example.

What's next

robrain outcomes feeds git reverts back into memory quality on both tiers; next is widening that to incidents and cycle time, so RoBrain can surface when a team is optimizing for the wrong thing in its own codebase.

Requirements

• Docker + Docker Compose (runs Postgres and Perception locally)
• Node.js with pnpm (build and CLI)
• An LLM key for the classifier — Anthropic Haiku or OpenAI
• An embedding key (e.g. OpenAI)
• No data leaves your machine in self-hosted mode

Docs

• Concepts (how it works, two pillars, Synthesis, comparisons) → docs/concepts.md
• CLI reference (explain, install, upgrading, editor setup, full command table) → docs/cli.md
• Maintainer release checklist (tag → GHCR → npm → MCP registry) → docs/release.md
• Troubleshooting (silent 401s, Docker rebuilds, stale summaries) → docs/troubleshooting.md
• Memory interchange format (robrain export, robrain-memory/v1 JSONL) → docs/memory-interchange.md
• VetoBench (does memory stop rejected re-proposals? methodology + archived receipts) → packages/vetobench/README.md
• Claude Code plugin (hook-based capture + veto warnings) → plugins/claude-code/README.md
• Codex CLI hooks (same lifecycle hooks, wired by robrain install) → plugins/codex/README.md
• Hermes agent plugin (memory-provider, capture + veto warnings) → integrations/hermes/robrain/README.md

Contributing

Apache 2.0. PRs welcome for extraction accuracy, new editor integrations, and embedding providers. See Concepts — Reference for tradeoffs and schema.

License

Apache 2.0 — see LICENSE

Built by Rory Plans