Most memory servers store and retrieve; this one returns a verdict. Checking a proposal against logged decisions and getting back "conflict" with the original reasoning and the alternatives you already rejected is what stops an agent quietly reopening a settled question three weeks later. The decision review inbox is the other half — it proposes and you confirm, so nothing archives itself.
A local-first memory server for the state a conversation log does not carry: what was decided and why, what is blocked, what is active, and what should not be reopened. The wedge is enforcement — before acting, an agent can check a proposal against your logged decisions and be told it is re-litigating a settled choice.
- Decisions with teeth — `decision_log` records the reasoning and the rejected alternatives, `decision_check` returns clear, caution or conflict on a proposal, `decision_refresh` updates review dates and evidence without touching the content, `decision_review` buckets overdue decisions and recommends an action for each
- Entities — `entity_read` and `entity_update` track projects and initiatives with status, momentum, blockers and next moves
- Plans — `plan_set` makes step 1 the active next move, `plan_advance` completes or skips a step and requires evidence or a reason, `plan_add` and `plan_read`
- Prioritization — `focus_get` recommends what to work on, and `project_evidence_scan` grounds it in the repository's own state: STATE.md, HANDOFF, ROADMAP, git activity and dirty files
- `context_resolve` — works out which entity the current work belongs to from mentions, aliases and files, deterministic and confidence-scored rather than guessed
- Hygiene — `pattern_detect` finds recurring blockers, stale work and avoidance signals, `memory_check` flags contradictions and noise, `memory_commit` saves at end of session, `audit_log` reads the full mutation history
- Safety — `risk_assess` classifies risky actions including destructive and public-release ones, `action_guard` applies guard templates before proceeding, `wrap_check` and `wrap_auto` catch accumulated state changes
- `semantic_recall` for searching memory by meaning, and slash commands installed into `.claude/commands/` so the agent has a fixed vocabulary
No account and no key for the core. Node.js 20 or newer; `npx brain-os init` in your project creates `.brain/`, installs the slash commands, and drops `AGENTS.md` plus thin pointer files for Claude Code, Copilot, Cursor, Zed and Windsurf — `--minimal` skips all but the first two. Connect with `claude mcp add brain-os -- npx brain-os serve`, or a `command`/`args` entry elsewhere. `semantic_recall` is the one tool that needs more: install `openai` alongside and set `BRAIN_EMBEDDINGS` to `openai`, which uses `text-embedding-3-small`. Reference `OPENAI_API_KEY` from your shell rather than pasting it into the MCP config, which is plaintext. Without the provider, that one tool returns a setup error and everything else keeps working. Package `brain-os`, MIT.
One command — npx brain-os init