On this page
← ML works

MemBridge

1 min read · Memory · FalkorDB · MCP · Retrieval

problem
An assistant's memory of you lives on someone else's servers, and you cannot read, correct or take it with you.
approach
Import ChatGPT or Claude history once, keep a provenance-linked, bi-temporal knowledge graph on your own machine (SQLite archive, FalkorDB graph, hybrid retrieval), and let MCP-capable assistants read and update it.
stack
Memory · FalkorDB · MCP · Retrieval
links
Code ↗

What it is#

MemBridge is a local-first personal AI memory. You import your ChatGPT or Claude history once, and it keeps a knowledge graph on your own machine, with every fact linked back to the message it came from. Any assistant that supports MCP can then read and update it. There is no mandatory cloud and no dependence on any provider's memory feature.

Why it exists#

I asked ChatGPT to dump everything it remembered about me and saw that it held my health, money and family in one paragraph that I had never been shown. I wanted to read that memory, correct it and take it with me. I built MemBridge for the FalkorDB hackathon. The write-up tells the story.

How it works#

An import goes into a raw SQLite archive. Rules, with an optional local model, extract memories and entities, remove duplicates, mark which fact supersedes which, and group related entities. The result is a FalkorDB graph that can be rebuilt from the archive. Retrieval combines keyword search, graph search, PageRank, rank fusion and a packing step, and is served over MCP or a localhost dashboard.

Health, money, identity-document and family facts are classified as private and hidden from every assistant by default. An assistant can see them only through a short, PIN-approved grant, and it can only request that something is forgotten; you confirm it. When there is no evidence, the answer is "I don't know".

Stack#

Python, SQLite, FalkorDB, MCP, a small web dashboard, Docker for the graph database.

Status#

On a 48-question multi-hop benchmark, text-only retrieval fails all 12 questions whose answers share no words with the query, while the graph-based system answers all of them. The repository also documents the weaknesses found. A read-only sandbox on synthetic data is hosted for trying it.