Mem0's local FAISS adapter deletes old memory before validating the replacement

izgorodin · reddit · 2026-11-04

A detailed engineering analysis of a Mem0 local FAISS adapter flaw: rejecting a bad memory update doesn't mean the old record survived.

The SDK reopened a 1024-dim collection with a 512-dim embedder. Updating raised an error, yet fetching the original record returned nothing — the adapter's update order is delete → save → insert, so removal happens before validation.
Patch 1 validates numeric conversion and dimensions before mutation; but insertion/save can still fail afterwards.
Patch 2 (Nov 3) prepares the replacement on a cloned index, saves, then publishes new in-memory state, plus a rollback journal for the FAISS index + JSON metadata pair. Both patches are pending issue acceptance, not merged.
Suggested regression check: save a known record, attempt an invalid replacement, then verify old content immediately and after reopening — asserting only that the call raised misses the state it left behind.
Scope: SQLite history retained the original ADD event; only the local FAISS adapter is affected, not Mem0 Cloud. The author notes this is an AI-assisted static reading of public sources; the v2 patch was implemented by an AI agent with 58 passing tests.

More from coding & agent