Way-audit of Mem0 OSS: 1 covered, 5 partial, 2 gaps (with file refs)

[alexk202]

Hi Mem0 team — we applied our memory audit checklist (MAQS-Way v0.2,
https://github.com/alexk202/maqs-way/releases/tag/v0.2) to Mem0 OSS
(audited commit 0724291) and wrote up the result as a worked example:

https://github.com/alexk202/maqs-way/blob/main/docs/audits/mem0-way-audit.md
Quick scope clarification to our post above (template norm we adopted after
publishing): this audit covers the OSS repo at commit 0724291; hosted
platform behavior is unjudged — "gap" means "absent in OSS", never "absent
at Mem0". Scale used: covered = proven in code, partial = mechanism exists
with one property missing, gap = absent. Corrections policy: wrong reference
→ we fix the pointer, finding stands; genuine misread → finding WITHDRAWN.

Short version, all with file:line evidence:

• COVERED: tenant isolation (pre-top-k filters, unspoofable identity keys) — exemplary.
• PARTIAL: provenance (no source ref/confidence on the record), aging
  (decay is platform-only), deletion (history stays queryable), write
  integrity (no multi-store transaction, a few silent paths), logging
  (SQLite history as closest to an audit trail).
• GAP: model-output trust typing, vendored eval harness.

Tone note: this is meant as help, not a roast — each gap comes with a
concrete per-file suggestion. Corrections policy (restated): wrong reference → we fix the pointer, the finding stands; genuine misread → WITHDRAWN. Corrections go in as a second commit.

No action needed on your side unless you want to comment. Thanks for the
open codebase — it made this audit possible.

[alexk202]

Update: we've published a reproduction recipe for the MAQS-Way eval runs (LongMemEval oracle, mem0 pinned at 0724291b, llama.cpp + Llama-3.1-8B). Would you try evals/REPRODUCE.md §8 on a clean machine? Failed or partial attempts are just as useful as successful ones: the deviation log is what we're after.

https://github.com/alexk202/maqs-way/blob/main/evals/REPRODUCE.md