deepddw

deepDDW — Memory & Knowledge Base for DeepSeek Harness, Reachable from Any Device on Your LAN

English · 简体中文

deepDDW gives DeepSeek Harness (DSH) the three pieces it's missing — memory, a knowledge base, and document search — and makes all of it reachable from any device on your LAN. No app install. No DSH source changes.

• 🧠 Memory — cross-session long-term memory, write/search via DSH's official MCP interface
• 📚 Knowledge base — ingest & search your docs (industry docs, SOPs, research notes) anytime
• 🌐 LAN-wide multi-device — deploy once; phones, laptops and tablets on the same network all use the same DSH workbench

Try it in 60 seconds:

npm i -g @deepseek-ai/dsh              # 1. install official DSH on the server
git clone https://github.com/ccch713/deepddw.git && cd deepddw
./install.sh --with-dsh                # 2. install deepDDW
./install.sh --port 8600               # 3. start
# 4. from any device on your LAN: open http://<server-ip>:8600/  (phone: scan the QR to auto-pair)

📸 Screenshot: phone / tablet accessing the workbench (coming soon).

Status: v0.3.0 · MIT · CI (pytest + ruff) ✅ · Multi-device on LAN + workspace isolation + backup/restore + TLS (opt-in) + LLM reflections & ranked memory search · Listed in awesome-deepseek-harness · Roadmap below

---

Why deepDDW?

Most DSH extensions give you memory alone. deepDDW is a complete workstation — memory + knowledge base + document search + LAN-wide multi-device access:

In one sentence: turn a "personal toy" into a tool a small team can actually use — fully packaged, easy to deploy, low maintenance, ready for small businesses (up to ~20 people) for their daily AI workflow.

deepDDW is built on our DDW AI HUB platform — validated in enterprise deployments and packaged into the DSH ecosystem as open source (MIT).

How It Works (Technical Path)

📱 Phone / 💻 Desktop / 📱 Tablet / 🖥️ Laptop — any device on the LAN
   │                  (browser access, no App install)
   ▼
deepDDW Gateway (one server on the LAN)
   ├─ /dsh/*   proxy → DSH engine (official UI, model config & chat untouched)
   ├─ /api/*   proxy → DSH RPC/API
   └─ /api/v1/*  deepDDW capabilities: Knowledge Base / Memory / Docs / LLM config
        │
        │  DSH official MCP client (streamable-http)
        ▼
deepDDW MCP tools (auto-invoked by the model)
   ├─ mcp__deepddw__ddw_kb_search          knowledge base search
   ├─ mcp__deepddw__ddw_memory_put         write memory
   ├─ mcp__deepddw__ddw_memory_search      search memory
   └─ mcp__deepddw__ddw_docs_portal_search document search

Integration = DSH standard MCP: DSH natively supports MCP clients; deepDDW exposes a standard streamable-http endpoint — zero intrusion, zero changes to DSH source. The UI, settings and model configuration all remain official.

Why Does It Work on Your LAN?

The official DSH listens on localhost only (for security), so phones/tablets cannot connect. deepDDW solves this with gateway proxying:

• DSH stays bound to localhost (official security design preserved)
• deepDDW gateway listens on the LAN, any device opens http://<server-ip>:8600/ to reach the original DSH workbench
• All data stays on your server — never leaves the LAN

Deploy once, the whole family/team can use it — a capability the official DSH does not provide.

Quick Start

# 1. Install DSH (official) on the server
npm i -g @deepseek-ai/dsh

# 2. Install deepDDW (packaged, one command)
git clone https://github.com/ccch713/deepddw.git
cd deepddw && ./install.sh --with-dsh

# 3. Start
./install.sh --port 8600

# 4. Open from any device on the LAN:
#    http://<server-ip>:8600/   → original DSH workbench
#    Phones/tablets: "Add to Home Screen" for an App-like experience

# 5. Add your API Key in DSH Settings → Models
# 6. In chat, ask the model to "search the knowledge base" or "remember ..."
#    → it auto-invokes the mcp__deepddw__* tools

Requirements: one ordinary computer/server (8 GB RAM minimum, 16 GB+ recommended), Python 3.11+, no GPU needed (LLM via cloud API or local Ollama).

Windows (standalone exe)

A Windows build is produced automatically by the windows-build workflow — download deepddw-windows.zip from the latest Actions → Artifacts:

# 1. Unzip anywhere (no Python/Node needed on the target machine)
deepddw-windows/deepddw.exe

# 2. Optional: point data/config at a custom location
set DDW_DATA_DIR=%USERPROFILE%\.deepddw
set DDW_ACCESS_TOKEN=<your-token>

# 3. Start (listens on 0.0.0.0:8500) → open http://<host>:8500/health
deepddw.exe

Upgrades are cheap: replace the whole folder with a newer zip — your data (under %USERPROFILE%\.deepddw) is untouched. See docs/windows-packaging.md for the full evaluation (PyInstaller one-dir + CI auto-build) and the manual build steps.

Multi-Device on LAN (0.2.0)

deepDDW is built for up to 20 devices on your LAN sharing one gateway:

• Device identity — each browser persists a device_id (localStorage) and can set a friendly name on the launcher; reconnects keep the same identity.
• Online status — devices register/heartbeat to the gateway; /api/v1/status (Token-protected) shows who is online, active WebSockets, request counts, DB size and version. The launcher renders a live status card for admins.
• Rate limiting — sliding-window per Token + per IP (default 60 req/min/token, global cap → 503 overload protection); configurable via config/deployment.yaml → security.rate_limit.* or DDW_RATE_LIMIT_* env.
• SQLite concurrency — WAL + busy_timeout=5000 + synchronous=NORMAL on every connection, plus a process-wide write lock for cross-table transactions (20 concurrent writers verified, no database is locked).

POST /api/v1/device/register    # register / rename this device (idempotent)
POST /api/v1/device/heartbeat   # keep-alive
GET  /api/v1/status             # status panel (token required)

Also in 0.2.0:

• Workspace isolation — devices pick a workspace (default shared); memory/logs and MCP memory tools are scoped per workspace, docs filtered by slug prefix. Legacy clients are unaffected.
• Session resume across devices — recent session summaries (up to 5) with a "continue" button: pick up a conversation on your phone where you left off on desktop.
• One-click backup / restore — backup via API (downloadable); restore validates the SQLite file (header + integrity check) and keeps a .pre-restore safety copy before replacing the main DB.
• Optional TLS — one-command self-signed cert (scripts/gen_self_signed_cert.sh, 1-year), enabled via security.tls.*; HTTP unchanged when off. For external access prefer a Caddy/Nginx reverse proxy (see docs/tls.md).
• Version / upgrade check — /api/v1/version reports latest_version / update_available (GitHub releases, 1h cache, offline-degraded); the launcher shows an upgrade banner.

Security & Privacy

Tech Stack & License

deepDDW itself: MIT License — free to use, modify, and commercially deploy; keep the copyright notice.

See NOTICE for full third-party attribution.

Ecosystem & Feedback

• Extend with official DSH plugins: deepDDW keeps DSH's native plugin mechanism intact. Install official plugins straight from the npm registry via the DSH official command — the only channel we recommend, to avoid supply-chain poisoning:

  dsh plugin --profile web add <npm-package>   # official npm registry only
  

  See SECURITY.md for our third-party plugin disclaimer and what deepDDW guarantees (memory & knowledge base only; no data theft/exploitation/sale).
• Knowledge distillation: use whatever distillation skill / workflow you prefer — the methodology is yours; deepDDW provides the complete pipeline "distilled output → searchable knowledge base → model-usable". More plugins and tools are on the way.
• Memory / knowledge migration: knowledge base uses standard SQLite; memory is organized by namespace/key/value — import from other agents or tools.
• Feedback: we'd love to hear how you use it; stronger open-source tools are coming in future releases.

Roadmap

Only items actually planned or already delivered are listed here.

Delivered:

• [x] Multi-device on LAN (0.2.0) — device identity/online registry, status panel, rate limiting, SQLite WAL concurrency (up to 20 devices)
• [x] Workspace isolation (P1-1) — per-workspace memory/knowledge scoping at the gateway (default shared, backward compatible)
• [x] Session resume across devices (P1-3) — "recent sessions" summaries (up to 5) with a continue button
• [x] Optional TLS (P1-2) — one-command self-signed cert; external access via Caddy/Nginx reverse proxy (see docs/tls.md)
• [x] Backup / restore API (P2-1) — one-click backup, downloadable; validated restore with .pre-restore safety copy
• [x] Load-test report (P2-2) — 5/10/20 devices: 0% errors, P95 ≤ 126 ms, no database is locked (see docs/load-report.md)
• [x] Version / upgrade check (P2-3) — /api/v1/version probes the latest release (1h cache); launcher shows upgrade banner
• [x] Docker one-click deployment — docker compose -f deepddw-compose.yml up -d --build (verified on a real macOS arm64 host: core + SearXNG containers up, health/MCP/chat end-to-end green)
• [x] Session → document auto-ingest — conversations saved to the knowledge base via ddw.docs.save / ddw.session.docs MCP tools + REST API, searchable and traceable per session
• [x] Vector search enhancement — hybrid retrieval (SQLite FTS5/LIKE + LanceDB, RRF fusion; optional, degrades to keyword-only when LanceDB is absent)
• [x] Windows packaging — PyInstaller one-dir build via the windows-build CI workflow, distributed as an Actions artifact, verified in release v0.1.0 (see docs/windows-packaging.md)
• [x] Reflection & consolidation (LLM polish) — daily reflection follows a style guide (auto/professional/casual), enforces a progress/issues/tomorrow structure and avoids repeating the previous day; consolidation skips logging when the LLM judges the conversation valueless
• [x] Memory search quality — results ranked by relevance score (hit-count × layer weight: user > notes > reflection > logs, plus freshness) instead of insertion order; keyword-expansion cache expiry tested

deepDDW — enterprise-grade capability, open-sourced for everyone.