nexusmemo 0.2.3

Local-first AI memory layer that gives LLMs persistent, structured memory

NexusMemo

Local-first AI memory layer that gives LLMs persistent, structured memory.

"We are building the memory operating system that turns AI from a temporary assistant into a long-term collaborator."

Problem

LLMs are stateless. They forget previous sessions, lose architectural context, repeat mistakes, and require repeated explanations. NexusMemo fixes this by building a persistent knowledge graph from your conversations.

How It Works

conversation → entity extraction → relationship graph → retrieval → prompt injection

When you tell NexusMemo something, it:

1. Extracts entities, relationships, and decisions via Claude (Client-side)
2. Embeds the text for semantic search
3. Builds a knowledge graph with NetworkX
4. Stores everything in a local SQLite database

When you query, it:

1. Searches semantically similar memories
2. Expands through the knowledge graph
3. Reranks by importance and relevance
4. Compresses into context ready for LLM injection

Quick Start

Install

pip install -e .

Configure

cp .env.example .env
# Optional configuration, but everything works locally by default

Use via CLI

# Add a memory
nexusmemo add "We replaced Airflow with Dagster because DAG maintenance became difficult"

# Search memories
nexusmemo search "What orchestration tool do we use?"

# Check status
nexusmemo status

# Get entity info
nexusmemo entity "Dagster"

Use via API

# Start the server
nexusmemo serve

# Add memory
curl -X POST http://localhost:8765/memory/add \
  -H "Content-Type: application/json" \
  -d '{"text": "We migrated from MongoDB to PostgreSQL for better ACID compliance"}'

# Query
curl -X POST http://localhost:8765/memory/query \
  -H "Content-Type: application/json" \
  -d '{"query": "What database do we use?"}'

Use with Claude Code (MCP)

The easiest way to run the MCP server is via uvx directly from PyPI.

# Add as MCP server (will automatically download and run the latest version)
claude mcp add nexusmemo -- uvx nexusmemo@latest mcp

Troubleshooting Path Issues: If Claude Code fails to find uvx (due to PATH environment issues), provide the absolute path to your uvx executable. You can find it by running which uvx in your terminal.

# Example with absolute path (replace with your actual path)
claude mcp add nexusmemo -- ~/.local/bin/uvx nexusmemo@latest mcp

Available MCP tools:

• add_memory — Store project context
• search_memory — Query past memories
• get_project_context — Get rich context about a topic
• find_related — Explore knowledge graph relationships
• get_decisions — Review past decisions
• memory_status — Check system stats

Architecture

┌──────────────┐
│   Interfaces │  CLI / FastAPI / MCP Server
├──────────────┤
│     Core     │  NexusMemo orchestration class
├──────────────┤
│   Services   │  Extraction, Embedding, Graph, Importance
├──────────────┤
│   Storage    │  SQLite + NetworkX (in-memory)
└──────────────┘

Tech Stack

• Python 3.11+ with FastAPI
• SQLite for persistent storage
• NetworkX for in-memory knowledge graph
• FastEmbed (ONNX) for local semantic search (Zero API Keys)
• MCP SDK for native Claude Code integration

Database

Four tables:

• memories — Raw text with embeddings
• nodes — Knowledge graph entities
• edges — Relationships between entities
• decisions — Project decisions with reasoning

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Run API server in dev mode
nexusmemo serve

License

MIT