vornicx/Midas

Overview

Midas is a local-first memory system for long-horizon AI agents that uses an eval-first approach and does not rely on an LLM during ingest. It provides a Python SDK and MCP server with source-traceable recall, belief revision, selective forgetting, and reproducible benchmarks.

Best for

Best for
Developers building long-horizon AI agents that need local, auditable memory with reproducible evaluation

Use cases

• Building agents that need persistent, traceable memory across long sessions
• Implementing selective forgetting or belief revision in agent workflows
• Running reproducible benchmarks for agent memory performance

How to use

Install

uv tool install "midas-memory[mcp,local]"

Tools exposed

• uv
• midas-memory

Tested with

Claude Code, Codex, Cursor, Claude Desktop, Windsurf, VS Code, Gemini CLI, Cline

Example client config

{\n  "mcp_servers": [\n    {\n      "url": "http://127.0.0.1:7077/mcp"\n    }\n  ]\n}

Notes

Midas is a local-first memory system for long-horizon AI agents that uses an eval-first approach and does not rely on an LLM during ingest. It provides a Python SDK and MCP server with source-traceable recall, belief revision, selective forgetting, and reproducible benchmarks.

12 stars on GitHub. Last updated 2026-07-12. Licensed Apache-2.0.

Use cases

• Building agents that need persistent, traceable memory across long sessions
• Implementing selective forgetting or belief revision in agent workflows
• Running reproducible benchmarks for agent memory performance

Pros

• Local-first design keeps data private and reduces latency
• Eval-first approach enables reproducible testing without LLM dependency at ingest
• Source-traceable recall improves debugging and auditability

Cons

• Small community with only 12 GitHub stars, limiting support and examples
• Requires Python environment, not language-agnostic
• No LLM at ingest may limit flexibility for complex semantic understanding

Indexed from awesome-mcp-servers-punkpeye and enriched against its public facts.