Skip to content

integrations.md

Latest commit

 

History

History
146 lines (115 loc) · 6.01 KB

integrations.md

File metadata and controls

146 lines (115 loc) · 6.01 KB
Raw

Integrations

Honest positioning. Every claim below is verifiable from the source tree, the public PyPI artifact, or a published benchmark file in benchmarks/. Nothing here is a customer-relationship claim about any AI vendor — the integrations described are software-level (the named tool talks to MIND-Mem via the Model Context Protocol), not commercial.

What MIND-Mem actually ships

Native MCP integration with 17 AI development tools

MIND-Mem speaks the Model Context Protocol. Any MCP-compatible client connects with one command:

pip install mind-mem
mm install-all

mm install-all auto-detects every supported client on your machine and writes the appropriate config file for each.

Currently supported:

Client Vendor Config
Claude Code Anthropic ~/.claude/settings.json
Claude Desktop Anthropic ~/Library/Application Support/Claude/
Codex CLI OpenAI ~/.codex/config.toml
Gemini CLI Google ~/.gemini/settings.json
Vibe (Mistral CLI) Mistral ~/.vibe/config.toml
Cursor Anysphere ~/.cursor/mcp.json
Windsurf Codeium ~/.codeium/windsurf/mcp_config.json
Zed Zed Industries ~/.config/zed/settings.json
Continue Continue.dev .continue/config.json
Cline Cline.bot VS Code extension settings
Roo Roo Code VS Code extension settings
GitHub Copilot GitHub / Microsoft VS Code extension settings
Cody Sourcegraph ~/.sourcegraph/cody.json
Qodo Qodo ~/.qodo/config.json
aider aider-chat .aider.conf.yml
OpenClaw OpenClaw ~/.openclaw/openclaw.json
NanoClaw / NemoClaw OpenClaw forks ~/.openclaw/openclaw.json

What this means: each of these tools can call MIND-Mem's 83 MCP tools (recall, propose_update, scan, hybrid_search, mic_convert_tool, mic_inspect_tool, etc.) the same way it calls any other MCP server.

What this does not mean: none of these vendors are commercial customers, paying users, partners, or have endorsed mind-mem. The integration is at the protocol layer — their software talks to our software. Compatibility is open and unilateral.

Open-source distribution

pip install mind-mem
  • License: Apache-2.0
  • PyPI: mind-mem
  • Source: star-ga/mind-mem
  • Local model: mind-mem:4b (fully trained) ships via Ollama — no cloud API required for the extraction model

Compatible with major LLM providers

MIND-Mem's recall pipeline is provider-agnostic. Tested against:

  • Anthropic Claude (3.5 Sonnet, 4.x family)
  • OpenAI GPT (4o, 5.4)
  • Google Gemini (2.0 Flash, 3.1 Pro)
  • Mistral Large
  • Local: Ollama, vLLM, llama.cpp endpoints

The "compatibility" claim is at the API contract level — the same MIND-Mem server returns the same answers regardless of which LLM is asking. We do not use any provider's commercial relationship as a positioning artefact.

Reproducible benchmarks

All numbers below are reproducible from benchmarks/ and the matching pipeline configs in the README.

Benchmark Score Methodology
NIAH (Needle In A Haystack) 250 / 250 (100%) Hybrid BM25 + BAAI/bge-large-en-v1.5 + RRF (k=60) + sqlite-vec. See benchmarks/NIAH.md.
LoCoMo (external LLM judge, 10-conv, 1986 questions) 73.8% Acc≥50, mean 70.5 BM25 + RM3 query expansion → top-18 evidence → observation compression → answer → judge. Full 10-conv benchmark.
LoCoMo (external LLM judge, conv-0, 199 questions, hybrid pipeline) 92.5% Acc≥50, mean 76.7 Hybrid: BM25 + Qwen3-Embedding-8B (4096d) → RRF fusion → top-18 → compression → answer → judge.
LoCoMo Adversarial subset 97.9% Acc≥50 Subset of the conv-0 hybrid run; tests retrieval against intentionally-misleading distractor turns.

Comparisons: published numbers for Mem0 (66.88), Zep (65.99), Letta (74.0), Memobase (75.8), LangMem (58.10) on the same benchmark. MIND-Mem surpasses Mem0 and Letta on the same 10-conv LoCoMo benchmark with zero cloud infrastructure and local-only retrieval — no graph DB, no vector DB service, no LLM in the retrieval loop unless the operator opts in.

Production usage at STARGA

MIND-Mem is the daily-driver memory layer across STARGA's active projects, including mind, mindlang.dev, mind-inference, and arch-mind. Used internally for cross-session recall, contradiction detection, and audit-grade rationale chains during agent-driven development.

This is a STARGA-internal usage statement — first-party, verifiable in our own commit history. We do not extrapolate it into a third-party "trusted by" claim.

What we do not claim

  • "OpenAI is a customer" — false. OpenAI runs its own memory systems. Codex CLI integration is software-level (MCP), not a commercial relationship.
  • "Microsoft is a customer" — false. Copilot integration is via the VS Code extension MCP surface, not a Microsoft Inc. commercial relationship.
  • "Anthropic is a customer" — false. Claude Code is built on the MCP spec; MIND-Mem implements that spec; Anthropic has not endorsed, contracted with, or partnered with STARGA.
  • "Used by N production teams outside STARGA" — we have no telemetry. PyPI download counts measure installs, not active use, and we do not turn install counts into traction claims.

If a future integration becomes a real commercial relationship (signed contract, NDA, paid pilot), it will appear in the press release and on this page — not before.

Surface allow-list (where this section may be reused)

Verbatim copy of the section above is approved for:

  • README "Integrations" section
  • mindlang.dev / MIND-Mem product pages
  • Investor decks and one-pagers
  • Cold outreach emails
  • LinkedIn / X marketing
  • Press releases

Do not paraphrase in a way that drops the "via MCP integration" qualifier next to vendor names. The qualifier is what keeps the claim defensible.