Open source · MIT · Runs on your machine

Memory that shows its work.

AxonDb is local graph memory and persistent memory for AI agents. It links the people, projects and tools your agent hears about, and every recall returns the path it followed to find the answer.

Get started
recall "Which project does Priya lead, and what does it use?"
Priya Atlas Redis "Priya leads Atlas." "Atlas uses Redis…" unrelated memory
Features

Similarity finds what's close.
Graphs find what's connected.

Vector search returns sentences that sound like your question. AxonDb also follows the links between them, so answers that span two or three facts still come back.

01

Fully local

SQLite on your disk. No account, no API key, no hosted service.

02

Explainable recall

The traversed path is part of every result, so you can see why a memory surfaced.

03

Semantic + graph

Embedding search finds the entry points. Two-hop expansion with decaying scores finds the rest.

04

Forgetting never deletes

Memories are closed with an end timestamp, so history stays intact.

05

CLI, MCP, viewer

Use it from the terminal, plug it into Claude Desktop, or watch the graph in a local web viewer.

06

Swappable storage

All database access goes through one interface, so SQLite can be replaced later.

07

Local viewer

The graph viewer is a small web server on 127.0.0.1.

AXONDB_PATH=/FULL/PATH/TO/memory.db python -m uvicorn AxonDb.api:app --port 8000
08

One SQLite file

Memories live in ~/.AxonDb/AxonDb.db, or any path you choose. Delete the file to remove everything.

Your machine Agent client Claude, CLI, browser AxonDb remember, recall SQLite file nodes, edges Local models Hugging Face first run only
How it works

Two verbs. One graph.

REMEMBER

Write a sentence, grow the graph.

  1. The sentence is embedded.
  2. A local model extracts entities and relations.
  3. Matching entities merge with existing ones.
  4. Memories link through shared entities and similarity.
RECALL

Ask a question, get the trail.

  1. The question is matched against memories and named entities.
  2. The search spreads two hops along edges, with scores decaying each step.
  3. You get the memories and the path used to reach them.
Benchmark

Biggest gains where facts must connect.

100 generated memories, 100 questions, one answer each. A hit means the right sentence is in the top 5. Same questions and same embedding model for every method.

Read this before quoting the numbers.
  • The data is templated and clean. Expect smaller gaps on real conversations.
  • 100 questions is a modest sample, and the data and the system share one author.
  • A hybrid vector + keyword baseline was not tested.
  • One-hop questions are a tie.
Quick start

Running in a minute.

# install
git clone https://github.com/yashogale30/AxonDb
cd AxonDb
python3 -m venv .venv
source .venv/bin/activate
pip install .

# remember and recall
AxonDb remember "Priya leads the Atlas project."
AxonDb remember "Atlas uses Redis for seat locking."
AxonDb recall "Which project does Priya lead?"
# claude_desktop_config.json
{
  "mcpServers": {
    "axondb": {
      "command": "/PATH/TO/.venv/bin/python",
      "args": ["-m", "AxonDb.mcp_server"],
      "env": { "AXONDB_PATH": "/PATH/TO/memory.db" }
    }
  }
}
# tools: remember, recall, forget

The first run downloads two small models from Hugging Face. After that everything runs locally. Write complete sentences with names, not pronouns, for the best entity linking.

Honest limits

What it doesn't do yet.

Privacy: memories stay in a local SQLite file and AxonDb sends nothing to a hosted service. To remove everything, delete the database file.