Mimir

The Well of Knowledge, Rewritten.

A zero-server, in-process agentic memory system for LLM agents.

📚 Prospectus | 🐍 PyPI | 📦 npm | 🏗️ Architecture | ⚡ Quickstart

Mimir is an open-source, in-process memory system for AI agents — combining dense vector search, bitemporal knowledge graphs, and autonomous tool-calling into a single Python process. Built on Zvec (Alibaba's battle-tested vector engine) and SQLite, it delivers production-grade semantic memory with temporal reasoning, zero infrastructure, and minimal setup.

💫 Features

• Bitemporal Memory: Facts are never deleted. Old records are time-capped, preserving full history. Ask "Where does John live?" and "Where did John live last month?" from the same dataset.
• In-Process Vector Search: Zvec 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. No cloud, no Docker, no external databases.
• Agent-Ready: LangGraph state graph with archive_memory and search_memory tools, ready to bind to any OpenAI-compatible LLM.
• REST API: Built-in FastAPI server for JS/TS and cross-language access via a single mimir-server command.
• Runs Anywhere: As an in-process library, Mimir runs wherever your code runs — notebooks, servers, CLI tools, or edge devices.

📦 Installation

Python

pip install mimir-memory

With Zvec (requires Python 3.10–3.12):

pip install mimir-memory[zvec]

With REST server:

pip install mimir-memory[server]

Node.js

npm install mimir-memory

Note: The npm package is a TypeScript SDK that connects to the Mimir REST server. Start the server first with mimir-server.

✅ Supported Platforms

• macOS (ARM64 — Apple Silicon optimized)
• Linux (x86_64, ARM64)

⚡ One-Minute Example

Python (in-process)

from mimir import MimirStorage, create_agent
from langchain_core.messages import HumanMessage

# Initialize storage (creates ./mimir_data with Zvec + SQLite)
storage = MimirStorage()

# Create a LangGraph agent with memory tools bound
graph = create_agent(storage, api_key="sk-...", base_url="https://...")

# Chat — the agent autonomously archives and retrieves memories
result = graph.invoke({"messages": [
    HumanMessage(content="My name is John and I live in Chennai.")
]})
print(result["messages"][-1].content)

Node.js / TypeScript

import { Mimir } from "mimir-memory";

const mimir = new Mimir(); // connects to http://localhost:8484

await mimir.archive({
  content: "John lives in London",
  source: "user",
  relation: "lives_in",
  target: "London",
  scope: "user",
});

const result = await mimir.search({ query: "Where does John live?" });
console.log(result);

🏗️ Architecture

┌──────────────────────────────────────────────────┐
│  L1 & L2: Tools (archive_memory, search_memory)  │
├──────────┬───────────────────────────────────────┤
│          │  L3: Bitemporal Knowledge Engine       │
│  Zvec    │  ┌──────────────────────────────────┐  │
│  (dense  │  │  SQLite (bitemporal_graph)        │  │
│  vectors)│  │  valid_from / valid_to timestamps │  │
│          │  └──────────────────────────────────┘  │
├──────────┴───────────────────────────────────────┤
│  L4: Procedural Optimizer (trajectory learning)   │
└──────────────────────────────────────────────────┘

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 and rule extraction

📊 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, CPU)
Agent Framework	LangGraph + LangChain Core
REST Server	FastAPI + Uvicorn
LLM	Any OpenAI-compatible endpoint

🤝 Community

• 📄 Full Prospectus — Architecture deep-dive, competitive analysis vs Letta/Mem0/Zep/LangMem, and Zvec integration details
• 🐛 Issues — Bug reports and feature requests

❤️ Contributing

We welcome contributions! Whether you're fixing a bug, adding a feature, or improving documentation — your help makes Mimir better for everyone.

📜 License

MIT