Qwen Memory

Long-term memory for AI agents, powered by Qwen on Alibaba Cloud and exposed over the Model Context Protocol (MCP). Any MCP-capable agent gains durable, cross-session memory that accumulates experience, retrieves what matters within a limited context window, and forgets what is outdated.

Hackathon track: Track 1 - MemoryAgent. License: MIT. Copyright (c) 2026 JHELY GLOBAL SL.

Repository: https://github.com/John-CEO-HQ/qwen-memory-mcp

This project is a learning experiment for the Qwen Cloud Hackathon: a standalone MCP server for agent memory on Alibaba Cloud.

Documentation

Guide	Purpose
docs/TESTING-GUIDE.md	Master index: testing and hackathon checklist
docs/CREDENTIALS-AND-SETUP.md	Accounts, API keys, regions, cost guardrails
docs/PHASE1-REMOTE-INTEGRATION.md	Live Qwen / DashScope tests from your machine
docs/PHASE2-DEPLOYMENT-TESTING.md	Alibaba deploy + verify deployed URL
docs/INSTALL.md	Full install, local run, Alibaba production deploy, troubleshooting
docs/JUDGE-TESTING.md	Instructions for hackathon judges
deploy/README.md	Alibaba ECS / Function Compute quick reference
AGENTS.md	Agent conventions and isolation contract

Why

Agents feel sharp inside a single conversation and amnesiac across sessions. This server gives an agent a managed memory layer that does four things well:

• Write - extract a durable memory (preference, fact, commitment, event), with a Qwen-derived summary, tags, importance (salience), and kind.
• Search - semantic retrieval ranked by similarity + salience + recency + reinforcement.
• Recall context - pack the most critical memories into a fixed token budget, ready to inject into a small context window.
• Forget - a maintenance pass that consolidates related memories with Qwen and lets stale, low-value memories decay away.

Architecture

flowchart LR
  agent["Any MCP client / agent"] -->|"MCP: write / search / recall / forget"| server["Qwen Memory MCP server (stdio or HTTP)"]
  server --> service["MemoryService"]
  service -->|"embeddings + reasoning"| qwen["Qwen on Alibaba Cloud Model Studio (DashScope)"]
  service -->|"persist + retrieve"| store["MemoryStore"]
  store --> file["File / in-memory (local + demo)"]
  store --> mysql["Alibaba Cloud RDS / PolarDB for MySQL (production)"]

Memory lifecycle:

flowchart TD
  w["memory_write"] --> analyze["Qwen analyze: summary, tags, salience, kind"]
  analyze --> embed1["Qwen embed (text-embedding-v3)"]
  embed1 --> active["active memory"]
  active --> s["memory_search / memory_recall_context"]
  s --> rank["rank: similarity + salience + recency + reinforcement"]
  rank --> pack["pack into token budget"]
  s -.reinforce.-> active
  active --> f["memory_forget"]
  f --> cluster["cluster by embedding similarity"]
  cluster --> consolidate["Qwen consolidate cluster -> canonical memory"]
  consolidate --> outdated["flag contradicted items -> forgotten"]
  active --> decay["decay score below threshold -> forgotten"]

Quick start

npm install

# Offline demo (no API key needed - deterministic local intelligence):
npm run demo

# Run the test suite:
npm test

# Run as an MCP server over stdio (for MCP Inspector / desktop clients):
npm run build && npm start

To use the real Qwen models, copy .env.example to .env and set QWEN_API_KEY (and optionally QWEN_BASE_URL for your region). Without a key, the server automatically falls back to the offline deterministic intelligence so it always runs.

MCP tools

Tool	Purpose	Key inputs
memory_write	Persist a durable memory	userId, content, sourceSession?, salience?
memory_search	Top-k semantic recall	userId, query, k?
memory_recall_context	Critical memories packed to a token budget	userId, query, tokenBudget
memory_forget	Consolidate + decay maintenance	userId

All memories are namespaced by userId, so one server can serve many agents.

Transports

• stdio (MCP_TRANSPORT=stdio, default) - launched as a child process by a local MCP client.
• Streamable HTTP (MCP_TRANSPORT=http) - stateless JSON-RPC at POST /mcp with optional Authorization: Bearer <MCP_AUTH_TOKEN>, plus GET /health. This is the shape used for cloud deployment and remote per-user MCP URLs.

Storage

• MEMORY_STORE=memory - in-process, ephemeral (tests/demo).
• MEMORY_STORE=file - single JSON file at MEMORY_FILE_PATH (local default).
• MEMORY_STORE=mysql - Alibaba Cloud RDS / PolarDB for MySQL (production); schema is created automatically. Vectors are stored as JSON and scored in the app; see src/memory/mysql-store.ts for the AnalyticDB-PG (pgvector) upgrade path.

Alibaba Cloud / Qwen

The only integration points with Alibaba Cloud are src/qwen.ts (DashScope embeddings + chat) and src/memory/mysql-store.ts (RDS/PolarDB). See docs/INSTALL.md for full production setup and deploy/README.md for a short Alibaba quick reference.

Configuration

See .env.example for all variables (Qwen models, store selection, transport, auth token, and forgetting/decay tuning).

Layout

qwen-memory-mcp/
  src/
    qwen.ts               # Alibaba Cloud / Qwen (DashScope) intelligence  [PROOF]
    fake-intelligence.ts  # offline deterministic intelligence (tests/demo)
    intelligence.ts       # picks Qwen vs fake
    config.ts             # env-driven config
    types.ts              # domain types
    memory/
      store.ts            # MemoryStore interface
      file-store.ts       # file / in-memory store
      mysql-store.ts      # Alibaba RDS / PolarDB store               [PROOF]
      create-store.ts     # store factory
      ranking.ts          # retrieval ranking + token-budget packing
      forgetting.ts       # clustering + consolidation + decay
      service.ts          # MemoryService (orchestration)
    server.ts             # MCP server + 4 tools
    transports/
      stdio.ts            # stdio transport
      http.ts             # streamable HTTP transport (stateless)
    index.ts              # entry point
  demo/cli.ts             # multi-session offline demo
  test/                   # vitest suite
  deploy/                 # Alibaba Cloud deployment docs
  Dockerfile

License

MIT License. Copyright (c) 2026 JHELY GLOBAL SL. See LICENSE.