BrainLayer

Your AI has amnesia. BrainLayer fixes that.

Every architecture decision, every debugging session, every preference you've expressed — gone between sessions. You repeat yourself constantly. Your agent rediscovers bugs it already fixed.

BrainLayer gives any MCP-compatible AI agent persistent memory across conversations. One SQLite file. No cloud. No Docker. Just pip install.

"What approach did I use for auth last month?"     →  brain_search
"Remember this decision for later"                 →  brain_store
"What was I working on yesterday?"                 →  brain_recall
"Ingest this meeting transcript"                   →  brain_digest
"What do we know about this person?"               →  brain_get_person

Quick Start

pip install brainlayer

Add to your MCP config (~/.claude.json for Claude Code):

{
  "mcpServers": {
    "brainlayer": {
      "command": "brainlayer-mcp"
    }
  }
}

That's it. Your agent now remembers everything.

Cursor (MCP settings):

{
  "mcpServers": {
    "brainlayer": {
      "command": "brainlayer-mcp"
    }
  }
}

Zed (settings.json):

{
  "context_servers": {
    "brainlayer": {
      "command": { "path": "brainlayer-mcp" }
    }
  }
}

VS Code (.vscode/mcp.json):

{
  "servers": {
    "brainlayer": {
      "command": "brainlayer-mcp"
    }
  }
}

MCP Tools (12)

Every tool includes ToolAnnotations so agents know which calls are safe to run without confirmation.

All 14 legacy brainlayer_* tool names still work as aliases.

Architecture

graph LR
    A["Claude Code / Cursor / Zed"] -->|MCP| B["BrainLayer<br/>12 tools"]
    B --> C["Hybrid Search<br/>vector + FTS5"]
    C --> D["SQLite + sqlite-vec<br/>single .db file"]
    B --> KG["Knowledge Graph<br/>entities + relations"]
    KG --> D
    E["JSONL conversations"] --> W["Real-time Watcher<br/>~1s latency"]
    W --> D
    I["BrainBar UI<br/>NSStatusItem + NSPopover"] -->|UDS /tmp/brainbar.sock| BB["BrainBarDaemon<br/>MCP + brain bus"]
    BB -->|MCP socket protocol| B

Everything runs locally. Cloud enrichment (Gemini/Groq) and Axiom telemetry are optional.

Why BrainLayer?

BrainBar — macOS Companion

Optional native Swift menu bar companion split into two launchd-managed processes:

flowchart LR
    UI["BrainBar<br/>LSUIElement UI"] -->|"watch-brain-bus + commands<br/>/tmp/brainbar.sock"| D["BrainBarDaemon<br/>headless MCP server"]
    D -->|"single writer queue + reads"| DB["SQLite WAL<br/>~/.local/share/brainlayer/brainlayer.db"]
    D -->|"helper subprocess IPC"| H["Hybrid search helper"]

BrainBarDaemon owns the MCP server, /tmp/brainbar.sock, the single-writer path, the watch-brain-bus stream, and helper subprocess lifecycle. BrainBar owns only the NSStatusItem, transient NSPopover, SwiftUI surfaces, hotkey routing, and a reconnecting socket subscriber. Killing the UI does not stop the daemon socket.

bash brain-bar/build-app.sh    # Build, sign, install LaunchAgent

The build script builds both BrainBar and BrainBarDaemon, embeds both binaries in BrainBar.app, then installs com.brainlayer.brainbar.plist and com.brainlayer.brainbar-daemon.plist with ProcessType=Interactive. It refuses non-canonical checkouts and dirty trees by default (#265) and stamps each bundle with GitCommit, GitDescribe, and BuildTimeUTC in Info.plist (#264) so a stale install is diagnosable in seconds.

Writer Arbitration

Background producers run with BRAINLAYER_ARBITRATED=1 and append writes to ~/.brainlayer/queue/; com.brainlayer.drain.plist drains that queue every 500ms as the single writer. Trigram FTS maintenance is explicit via brainlayer repair-fts and the weekly com.brainlayer.repair-fts.plist, not synchronous startup work. See docs/arbitration.md.

Recent Hardening (2026-04-15 → 2026-05-17)

Two-week stability sprint behind the next presentation. Every line below traces to a merged PR.

Search recall & dedup

• FTS recall hardened across Python, Swift BrainBar, and the watcher pipeline (#263).

• Lexical defense dictionary exports for fragile-token recovery (#262).

• MMR post-retrieval dedup on brain_search (#242).

• Legacy unique content_hash index dropped — was blocking re-enrichment writes (#245).

• Swift brain_store queue fallback so BrainBar can persist when the daemon is mid-restart (#261).

BrainBar reliability & UX

• MenuBarExtra(.window) rewrite with live-state sparklines and full-width hero (#248).

• Dashboard UX overhaul (#246).

• MCP initialize handshake preserved under backpressure (#247).

• KG force-sim early-exit + onAppear timer reset — kills CPU pegging when the graph tab is idle (#249).

Phase B preventive infra (2026-05-01) — one canonical artifact per environment

• /post-merge-deploy-check skill + initial canonical-deploy-registry.json (orchestrator#60) cross-checks GitHub merge metadata, the registry, and the deployed app's Info.plist so a merged PR cannot be declared shipped while the local bundle still points at the wrong build.

• Canonical app paths corrected in the deploy registry schema (orchestrator#58).

• Build-stamp + canonical-build guards land together so future BrainBar bundles carry provenance and refuse silent worktree overwrites (#264, #265).

Test gates — pre-push gate is mandatory before any push to main

• Pre-push regression gate (#257) plus exit-0 fix on the success path (#260).

• scripts/run_tests.sh orchestrator unifies Python + Swift + isolation test runs (#256).

• Stale-index regression fixture (#255) and Deepchecks regression harness (#259).

Security

• All 11 Swift MCPRouter tools exposed via BrainBar now ship ToolAnnotations (cyberMaster H1) (#253).

In flight (2026-05-02 reliability sprint) — PR #251

• Restores the resizable dashboard panel via a floating NSPanel (BrainBarDashboardPanelController) instead of MenuBarExtra(.window).

• Adds trigram FTS5 (chunks_fts_trigram) with a startup-safety guard: synchronous backfill is skipped when the desynced trigram table exceeds 10K chunks, so BrainBar never blocks the live ~360K-chunk database before /tmp/brainbar.sock opens.

• KG atlas presentation (importance-based altitude filtering, region backdrops, deterministic seeding) and AgentActivityMonitor for live CLI presence on the dashboard.

• Pub/sub plane on /tmp/brainbar.sock is explicitly preserved (brain_subscribe, brain_unsubscribe, notifications/claude/channel) — only search/store handlers move to the Python MCP path.

Phase 5 ship wave (2026-05-17) — ingest hygiene + KG regression fix

• Diagnostic + PreCompact noise rejection at ingest (#289) — recursive_mcp_output_reason now detects BrainLayer-MCP-unavailable diagnostics and PreCompact checkpoint payloads, rejecting them at the watcher / drain / store ingestion heads so tooling failures do not become durable memory. The hybrid reranker demotes (not removes) any chunk tagged with precompact/quarantine signals so explicit include_checkpoints callers still see them. Pre-push gate: 1995 passed, 9 skipped, 75 deselected, 1 xfailed. A dry-run-first scripts/quarantine_noise.py is available for back-filling existing infra noise — live DB mutation requires explicit --apply.

• Persist digest LLM entities (#290) — fixes a KG persistence regression where brain_digest silently skipped Gemini entity extraction because process_chunk passed use_llm=llm_caller is not None and the MCP/CLI path never sets llm_caller. Non-seed person entities were never materialized into kg_entities / kg_entity_chunks. The 2026-04-06 entity-recall recurrence root-caused to this code path. RED-first regression test (test_digest_content_persists_llm_people_entities_for_lookup) now guards the fix.

• Enrichment LaunchAgent recovered — com.brainlayer.enrichment was silently unloaded since 2026-05-15 11:50 IDT (no entity extraction running). Bootstrapped back on 2026-05-17 against the 56K-chunk backfill; throttled by Gemini 503s on flex tier but actively draining (verified via launchctl list | grep enrichment returning a live PID).

Data Sources

Enrichment

Each chunk gets 10 structured metadata fields from a local or cloud LLM:

brainlayer enrich                    # Run enrichment on new chunks
BRAINLAYER_ENRICH_BACKEND=groq brainlayer enrich   # Force Groq

CLI Reference

brainlayer init               # Interactive setup wizard
brainlayer index              # Batch index conversations
brainlayer watch              # Real-time watcher (persistent, ~1s)
brainlayer search "query"     # Semantic + keyword search
brainlayer enrich             # LLM enrichment on new chunks
brainlayer stats              # Database statistics
brainlayer brain-export       # Brain graph JSON for visualization
brainlayer export-obsidian    # Export to Obsidian vault
brainlayer dashboard          # Interactive TUI

Testing

pip install -e ".[dev]"
git config core.hooksPath .githooks     # install repo pre-push hook once per clone
pytest tests/                           # 1,848 Python tests
pytest tests/ -m "not integration"      # Unit tests only (fast)
ruff check src/ && ruff format src/     # Lint + format
# BrainBar: 54 Swift tests via Xcode

See full configuration reference for all options.

pip install "brainlayer[brain]"       # Brain graph visualization + FAISS
pip install "brainlayer[cloud]"       # Gemini Batch API enrichment
pip install "brainlayer[youtube]"     # YouTube transcript indexing
pip install "brainlayer[ast]"         # AST-aware code chunking (tree-sitter)
pip install "brainlayer[kg]"          # GliNER entity extraction (209M params)
pip install "brainlayer[telemetry]"   # Axiom observability
pip install "brainlayer[dev]"         # Development: pytest, ruff

Contributing

Contributions welcome! See CONTRIBUTING.md for dev setup, testing, and PR guidelines.

License

Apache 2.0 — see LICENSE.

Part of Golems

BrainLayer is part of the Golems MCP agent ecosystem:

• BrainLayer — Persistent memory (this repo)

• VoiceLayer — Voice I/O for AI agents

• cmuxLayer — Terminal orchestration for AI agents

Originally developed as "Zikaron" (Hebrew: memory). Extracted into a standalone project because every developer deserves persistent AI memory.