Memory + Sequential Thinking with Qwen3 8B

Install + handshake last tested 2026-10-22 · Local-model tests: Not yet tested

Client: Any local MCP host (LM Studio shown). Client versions and features come from vendor docs, checked 2026-10-22. We haven't run this setup end-to-end yet.

Prerequisites

• A Qwen3 8B build in your local runner (Qwen3 8B itself: not yet tested by us; our runs used Qwen2.5 3B/1.5B)
• Node.js 18+ for npx
• Works fully offline once packages are installed (offline behavior: not yet tested)

Steps

1. Add both servers to your host's mcp.json (config below).
2. Start with the suggested tools only; add more if the model copes.

Config

{
  "mcpServers": {
    "memory-knowledge-graph-reference": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-memory"
      ]
    },
    "sequential-thinking-reference": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-sequential-thinking"
      ]
    }
  }
}

Suggested tool allowlist

Picked from the real tool names in our test; whether a small model succeeds with them is not yet tested.

• Memory / knowledge graph (reference) (9 tools listed): create_entities, add_observations, search_nodes, open_nodes
• Sequential Thinking (reference) (1 tools listed): sequentialthinking

Measured context cost

• Memory / knowledge graph (reference): ~925 tokens (11% of 8k, 3% of 32k) for all 9 tools. Small-model friendly: high.
• Sequential Thinking (reference): ~861 tokens (11% of 8k, 3% of 32k) for all 1 tools. Small-model friendly: medium.

Method: tiktoken cl100k_base over compact JSON of name+description+inputSchema (approximate). Cost for the allowlist subset alone: not yet measured.

Success rates on small models

• Memory / knowledge graph (reference): qwen2.5-3b-instruct-q4km: 100% tasks completed, 100% correct tool, 3 runs; qwen2.5-1.5b-instruct-q4km: 67% tasks completed, 100% correct tool, 3 runs (8 vCPU, 15 GB RAM, no GPU; context 16384; client: testedmcp smallmodel.mjs (Ollama /api/chat tools) + Ollama 0.40.8) Exploratory single 7B run (qwen2.5-7b-instruct-q4km): passed.
• Sequential Thinking (reference): Not yet tested

These were our own task runs, with all tools exposed. The setup on this page (client, model, allowlist) hasn't been run end-to-end.

Known failure modes (community-reported, not measured by us)

• Too many tools: the model picks the wrong one. Fix: use the allowlist above.
• Gemma thinking mode suppresses tool calls. Fix: turn thinking off.
• Context overflow from large tool schemas. Fix: fewer tools, or a bigger context window.

FAQ

Does Memory / knowledge graph (reference) / Sequential Thinking (reference) work with local models?

In our small-model runs (not this exact setup): Memory / knowledge graph (reference): qwen2.5-3b-instruct-q4km completed 100% of 3 runs; qwen2.5-1.5b-instruct-q4km completed 67% of 3 runs. Sequential Thinking (reference): not yet tested on local models.

How many tools does it expose?

Memory / knowledge graph (reference): 9 tools in our tools/list test; Sequential Thinking (reference): 1 tools in our tools/list test.

Why limit the tools?

Community reports say small models choose the wrong tool or emit malformed arguments when shown many tools. We haven't measured this yet.

Are there affiliate links on this page?

No.