Claude MarketplacesSubmit
vornicx avatar

Midas

vornicx/midas

A memory layer for AI agents that runs entirely offline and skips the usual LLM-based fact extraction step. Instead of sending every conversation turn through an API to summarize it, Midas uses local embeddings to decide what matters, retrieve relevant context by meaning, and trace every recalled memory back to its source turn. The MCP server exposes tools for storing facts, querying by semantic similarity, and managing belief revision without per-message API costs. Designed for long-running coding agents and assistants where you want durable memory across sessions but don't want to pay for or trust a third party to process your entire conversation history. Everything lives in a local SQLite file with a local embedding model, so there's no network dependency after the initial model download.

GitHub stars
6
tools
2
transport
STDIO
env vars
4
last updated
Jun 7, 2026
View on GitHub

Tools

2 tools

Public tool metadata for what this MCP can expose to an agent.

memory_stateReturns the live, non-superseded decisions, constraints, and facts.

Returns the live, non-superseded decisions, constraints, and facts.

No parameter schema in public metadata yet.

memory_diffLists beliefs added since a prior point and beliefs revised from old to new.

Lists beliefs added since a prior point and beliefs revised from old to new.

No parameter schema in public metadata yet.

Sponsors

Promote your product hereReach 125k+ visitors monthlyPromote your product hereReach 125k+ visitors monthlyAgent CaffeineSponsoredClose your lid, come back to finished workSnugSponsoredThe clipboard manager for the AI eraPromote your product hereReach 125k+ visitors monthlyWaxSponsoredBlock anything with supervision, without erasing your iPhonePromote your product hereReach 125k+ visitors monthlyknobsSponsoredSee your site on any phone, any day, in any language

Configuration

MIDAS_MCP_DB

Path to a SQLite file to persist memory across restarts (default: in-memory)

MIDAS_MCP_EMBEDDERdefault: local

Embedding backend: 'local' (bge ONNX, offline) or 'hashing' (default: local)

MIDAS_MCP_MAX_RECORDS

Cap the store; above it the lowest-value memories are auto-forgotten (no LLM)

MIDAS_MCP_MIN_IMPORTANCEdefault: 2

Relevance floor 1-5 for auto-capture; turns scoring below it are skipped

Readme

Midas

The local memory layer for long-horizon AI agents — remembers across sessions, keeps what's current, and won't act on stale memory.
No LLM at ingest · $0 per message · fully local · every recall traces to its source.

tests PyPI npm Python 3.11+ License: MIT

View the full README on GitHub

Details

Repository
github.com/vornicx/midas
Category
AI & LLM Tools
Package
midas-memory-mcp
Transport
STDIO
Registry
MCP registry, active

Related AI & LLM Tools MCP servers

View all
1bahfahh avatarNoteit Mcpbahfahh/noteit-mcpMCP server for AI agent profiles and smart notes. 60+ coding prompt packs with expert personas.5stars
2blackfoil avatarClaude Token Saver Mcpblackfoil/claude-token-saver-mcpMCP server that offloads routine coding tasks to local LLM (Ollama) to save Claude API tokens4stars
3j-east avatarPixel Surgeonj-east/pixel-surgeon-mcpAI image and video generation, editing, and region repair via Gemini, OpenAI, and Grok.4stars
4lore-context avatarLore Contextlore-context/lore-contextGoverned AI-agent memory, Evidence Ledger traces, evals, and portable context tools.4stars
5n24q02m avatarImagine Mcpn24q02m/imagine-mcpMCP server for image/video understanding & generation (Gemini/OpenAI/Grok)4stars
6madezmedia avatarAcmi Mcpmadezmedia/acmiPersistent agent memory via ACMI — Profile/Signals/Timeline. 16 tools on Upstash Redis.3stars