Skip to content

README.md

Latest commit

 

History

History
117 lines (91 loc) · 4.6 KB

README.md

File metadata and controls

117 lines (91 loc) · 4.6 KB
Raw

Memex — server variant (single Docker image)

A server-based Memex that runs Qdrant + the web service + MCP in one container — no macOS app required. It's additive: the Tauri desktop app is untouched; this is a second way to run Memex, aimed at the Claude CLI and browsers.

Surface Where Port
Web UI (static) / 8765
JSON API /api/* 8765
MCP over HTTP /mcp (POST JSON-RPC) 8765
Qdrant (inside the image) REST /readyz + dashboard / gRPC 6333 / 6334

Build & run (one image)

# from the repo root
docker build -t memex-allinone -f deploy/web/Dockerfile .
docker run --rm --name memex-allinone -p 8765:8765 -p 6333:6333 memex-allinone

That single container starts Qdrant internally, waits for /readyz, then serves the web service. Persist the indexed corpus by mounting a volume at /qdrant/storage:

docker run --rm --name memex-allinone -p 8765:8765 -v memex_qdrant:/qdrant/storage memex-allinone

Load the sample corpus

The image ships the synthetic corpus at /app/sample-corpus:

curl -fsS -X POST http://localhost:8765/api/index \
  -H 'Content-Type: application/json' -d '{"path":"/app/sample-corpus"}'
# → {"indexed":12,"total":12,"errors":0,...}

The embedding model is pre-baked into the image (no first-query network needed), and the container auto-indexes the bundled corpus on startup (MEMEX_WEB_AUTOINDEX=1) — so the browser UI shows data immediately. The manual /api/index call above is only for (re)indexing a different corpus. Point at your own sessions by mounting them under /home/memex/.claude/projects and setting MEMEX_SCAN_ROOT.

JSON API

curl -fsS http://localhost:8765/api/health
curl -fsS "http://localhost:8765/api/search?q=rate%20limiter%20redis&limit=3"
curl -fsS "http://localhost:8765/api/recall?q=cargo%20build%20linker%20error"
curl -fsS "http://localhost:8765/api/topology?sample=12&per_point=4"
curl -fsS -X POST http://localhost:8765/api/lens -H 'Content-Type: application/json' \
  -d '{"query":"build error","weights":{"error":2.0,"content":1.0}}'
curl -fsS -X POST http://localhost:8765/api/mix -H 'Content-Type: application/json' \
  -d '{"pos":["<session_id>"],"neg":["<session_id>"]}'

Register with the Claude CLI (both transports)

HTTP (remote MCP — recommended for the web service):

claude mcp add --transport http memex-web http://localhost:8765/mcp

stdio (the same binary, inside the running container):

claude mcp add memex-web -- docker exec -i memex-allinone memex mcp

Both expose the same 9 tools (find_similar_sessions, find_similar_error, predict_next_action, mix_similar_sessions, get_session_summary, get_session_turn, list_recent_sessions, analyze_corpus_topology, snapshot_export).

Local dev (no Docker)

bash scripts/start-qdrant.sh                       # Qdrant on :6333/:6334
cargo run --release --manifest-path src-tauri/Cargo.toml \
  --no-default-features --features web -- serve --port 8765 --ui-dir src

How it's built (additive, Tauri-free)

The web binary compiles with --no-default-features --features web: the tauri/WebKit dependencies are gated behind the default gui feature, so this build links neither and runs on debian:trixie-slim. The macOS app build (cargo build --release, default features) is unchanged.

Browser UI

The served UI is functional in a browser: a __TAURI__ fetch shim (injected only on the web-served index.html) routes the frontend's invoke("cmd", args) calls to POST /api/invoke/{cmd}, which dispatches to the same backend logic the desktop app uses. The Tauri desktop app is unaffected (it loads index.html directly, with the real Tauri runtime).

Hardening & self-containment

  • Non-root: runs as user memex (uid 10001); ports are all > 1024.
  • Proper init: tini is PID 1 (signal handling + zombie reaping) over the bash entrypoint that supervises Qdrant + the web server.
  • Self-contained: the BGE-small model is pre-baked at build time — no network needed at runtime.
  • Replay & predict work: the corpus is mounted under /home/memex/.claude/projects, a trusted sec-sandbox root, so predict and turn-by-turn replay (which re-parse source .jsonl) function in-container.

Remaining notes

  • The bundled corpus is synthetic sample data; mount your own under /home/memex/.claude/projects (+ MEMEX_SCAN_ROOT) for real sessions.
  • tail_recent_errors (proactive-recall polling) returns empty in the server variant — a static server corpus has no live sessions changing under it.