Native Swift app + MCP server. Your notes, your device, zero cloud. Works with Claude, Hermes, OpenClaw. Install in 30 seconds.
BMem is a memory provider for Hermes, Claude and OpenClaw — a native macOS app + MCP server for semantic memory over your Markdown notes. A Basic Memory alternative that runs entirely on your Mac: on-device embeddings on the Neural Engine, a Spotlight-style hotkey search panel, and an OpenAI-compatible embeddings endpoint. No Docker, Qdrant or Redis — one app.
⬇ Download the latest release current build: BMem-1.1.2-beta3.dmg (442 MB, model included) — requires macOS 27+ on Apple Silicon.
- 🔒 Privacy First — no telemetry, no cloud, no external APIs. Your data never leaves your Mac. The telemetry toggle has no network code behind it at all: it is a no-op stub until signed builds exist.
- ⚡ Native Performance — the daemon idles at ~30 MB RSS / 77 MB footprint; model weights unload after 335 s idle and the embed worker exits after 2×335 s. Warm searches answer in well under a second (p90 0.32 s on an 11.1k-chunk corpus, measured 2026-10-19).
- 🎯 Zero Setup — the dmg ships the embedding model inside the app. No Python, no uv, no Docker, no API keys, no terminal.
- 🔄 Seamless Migration — same Markdown files as Basic Memory. Your
existing Basic Memory config is picked up automatically on first launch
(seeded from
~/.basic-memory/config.json); switching your agent is one config line (basic-memory→bmem mcp), and back just as easily. No lock-in in either direction — Obsidian reads the same files.
# 1. Download the latest DMG (button above)
# 2. Open it, drag BMem.app to Applications
# 3. Launch BMem, pick your notes folder
# 4. Register BMem with your AI agent — for Claude Desktop
# (claude_desktop_config.json):
{
"mcpServers": {
"bmem": {
"command": "/Applications/BMem.app/Contents/MacOS/bmem",
"args": ["mcp"]
}
}
}
# Done — your agent now has persistent memory.Use the full
/Applications/BMem.app/Contents/MacOS/bmempath: Claude Desktop is a GUI process and does not inherit your shellPATH, so a bare"bmem"will not resolve. Your notes folder is chosen in the app window on first launch; power users can override it with theBMEM_PROJECT_PATHenvironment variable.
The same command works for every stdio MCP client — Hermes
(mcp_servers block in config.yaml), OpenClaw, Claude Code
(claude mcp add bmem -- /Applications/BMem.app/Contents/MacOS/bmem mcp),
Codex, Cursor, VS Code. Familiar tools stay — no skills or prompts to
rewrite. Verified on 1.1.2-beta3: the handshake above returns
14 tools (search_notes, read_note, build_context, …).
Measured on 2026-10-19: dmg mounted → app in /Applications → launched → first search answered in 16 seconds (reinstall over an existing index; a first-ever launch additionally indexes your vault once).
The model lives in a child embed-worker that is killed when idle; the daemon keeps serving keyword results meanwhile. Real run on the live daemon (1.1.2-beta3, 11.3k chunks, 2026-10-19), footprint sampled every 15 s:
footprint MB D = daemon W = embed worker (model resident)
400 │
343 │
286 │
258 │ WW WW WW WW WW WW WW WW WW W
229 │ W WW WW
200 │W WW WW W
172 │ W WW W
115 │D DD DD DD DD D
86 │ D DD DD DD DD DD DD DD DD DD DD DD D
└─────────────────────────────┼────────────────────────
t=0 cold query 6.6s 335s idle → model unload
(worker exits after 2×335 s idle)
After the worker exits the steady idle is 77 MB footprint / ~30 MB RSS. A cold query after unload costs ~6 s (worker respawn + model load; ~60 s once after an OS update while the ANE shader cache rebuilds). Worker RSS looks bigger (4–7 GB) because CoreML pages count compressed memory; footprint is the honest number and is what the graph shows.
70 real RU/EN reference questions on a live personal-notes corpus (11,147 chunks), frozen internal suite — not an industry benchmark. Fresh A/B run on 1.1.2-beta3, 2026-10-19:
| Metric | Qwen3-Embedding-0.6B | EmbeddingGemma-2 (default) |
|---|---|---|
| top-1 | 46/70 (65.7 %) | 56/70 (80.0 %) |
| top-3 | 62/70 | 64/70 |
| top-1 RU / EN | 30/44 · 16/26 | 37/44 · 19/26 |
| search median | 348 ms | 270 ms |
| search p90 | 559 ms | 324 ms |
| idle footprint after unload | 182 MB | 89 MB |
Method, caveats and the anonymized query set: evals/BENCHMARKS.md.
Star counts pulled from the GitHub API on 2026-10-19.
Direct alternatives (agent memory over your notes):
| Project | Stack | Stars | Notes |
|---|---|---|---|
| basic-memory | Python + uv | 4.3k | canonical note format — BMem is file-compatible with it and auto-imports its config |
| mem0 | Python | 67k | memory-layer API/library, cloud-first with OSS tier |
| mcp-memory-service | Python + ChromaDB | 2.0k | MCP server, needs a vector store |
| OpenMemory MCP | mem0 stack + Docker | — | local memory UI on the mem0 engine |
| mcp-local-memory | TypeScript, SQLite FTS5 | 60 | keyword-first, optional sqlite-vec |
Adjacent (not file-based note memory): Graphiti/Zep (32k), Cognee (32.5k), Letta (25.2k) — enterprise agent-memory stacks; Supermemory (31.3k) — cloud API; agentmemory (29.7k) — coding-agent session capture daemon.
What none of them do in one native app:
- native Swift — no Python, no Docker, no Redis, no Qdrant, no ChromaDB;
- embeddings on the Apple Neural Engine via CoreML (not GPU, not a server);
- a Spotlight-style global hotkey panel for humans, not just agents;
- file compatibility with Basic Memory + automatic import of its config;
- an OpenAI-compatible
/v1/embeddingsendpoint on your Mac; - LAN sharing of the same index behind an API key;
- parallel work with Obsidian on the same files.
- macOS 27+ only — older macOS is not supported and won't be soon.
- Apple Silicon only — no Intel builds (the ANE path is the product).
- It is agent memory, not knowledge management — if you want a PKM/Zettelkasten suite for humans, that is Basic Memory's or Obsidian's job; BMem indexes the same files for your AI agents.
The roadmap isn't published yet — but the issue tracker is. Tell us what you're missing; it may well be the next thing we build.
The dmg contains EmbeddingGemma-2 inside the app bundle — nothing to download.
If you build from git, place the models manually into ~/.bmem/models/:
- Default engine: FluidInference/embeddinggemma-2-coreml (CoreML conversion of Google's EmbeddingGemma-2 weights, runs on the ANE)
- Fallback engine: coreai-community/Qwen3-Embedding-0.6B-CoreAI (CoreAI conversion of Qwen's Qwen3-Embedding-0.6B weights)
| Search | Hotkey panel | Memory & rollbacks |
|---|---|---|
 |
 |
 |
More (all English, shot on 1.1.2-beta3): first launch · what BMem reads · permissions · advanced · system · Russian set
- Basic Memory — the upstream ecosystem whose note format and MCP surface we are compatible with. We respect the upstream and chose AGPL deliberately. Also: basicmemory.com
- FluidInference/embeddinggemma-2-coreml — default embedding engine (CoreML build; original weights by Google)
- coreai-community/Qwen3-Embedding-0.6B-CoreAI — fallback engine (CoreAI build; original weights by Qwen)
AGPL-3.0-only — see LICENSE. Want to embed BMem in a commercial product? AGPL obligations still apply — for a commercial license contact the author: open an Issue or use the email in the GitHub profile.
💛 Want to support BMem? We're not taking money through payment rails yet — but if you'd like to sponsor development (or commercial licensing), open an issue and we'll sort out the details.