mimir-memory 0.2.0

Zero-server, in-process agentic memory system for LLM agents — Zvec + SQLite + LangGraph

Mimir — Local Agentic Memory System

A zero-server, in-process memory system for LLM agents, optimized for Apple Silicon.

Mimir combines dense vector search (Zvec), bitemporal knowledge graphs (SQLite), and autonomous tool-calling (LangGraph) into a single Python process. No Docker, no cloud databases, no external servers.

✨ Key Features

• Bitemporal Memory — Facts are never deleted. Old records are time-capped (valid_to), preserving full history. Ask "Where does John live?" and "Where did John live last month?"
• In-Process Vector Search — Zvec (Alibaba Proxima) runs as a C-extension inside your Python process. Sub-millisecond search, zero network hops.
• Fully Local — Embeddings via HuggingFace BGE run on CPU. Storage is file-based. The only external call is to your LLM endpoint.
• LangGraph Agent — Tools (archive_memory, search_memory) are bound to the LLM and routed via a compiled state graph.

🏗️ Architecture

Layer	Component	Role
L1 & L2	Tools	archive_memory and search_memory — the agent's interface to storage
L3	Storage	Zvec (dense vectors) + SQLite (bitemporal graph edges)
L4	Optimizer	Post-session trajectory analysis (stub)

🚀 Quick Start

# Requires Python 3.10-3.12 (for Zvec binary compatibility)
# Install Python 3.12 via pyenv if needed:
pyenv install 3.12.8
~/.pyenv/versions/3.12.8/bin/python3 -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Run
python3 mimir.py

📦 Stack

Component	Technology
Vector Storage	Zvec v0.2.0 (Alibaba Proxima)
Graph/Relational	SQLite3 (Python stdlib)
Embeddings	BAAI/bge-small-en-v1.5 (384-dim, local)
Agent Framework	LangGraph + LangChain Core
LLM	Any OpenAI-compatible endpoint

📄 Documentation

• PROSPECTUS.md — Full architecture deep-dive, competitive analysis vs Letta/Mem0/Zep/LangMem, and Zvec integration details

📜 License

MIT