Repository navigation
Replies: 2 comments
|
Your reading of the docs is accurate — the dedicated Memory Expiration page (added July 2026) and the add/update API references all say the same thing: On your three questions, from what's documented:
Happy to look at your boundary test cases if you share them — exact-instant and leap-day cases are the ones where day-granularity systems usually surprise people. |
|
Hi @HUIIIM — following up on my comment above: were you able to try it? If it solved the issue, marking it as the answer helps the next person who hits the same thing. If not, tell me what you saw and I'll take another look. |
Uh oh!
There was an error while loading. Please reload this page.
Hi mem0 team — we're building a personal AI memory layer (N=1, single user) and have been studying how production memory systems handle expiring facts. When we first drafted this post, mem0 had no documented expiry semantics we could find; the new
expiration_datefeature answers that first question, so this post is now only about the boundary choices, where we landed differently and would value your reasoning:Your semantics (from the docs, read 2026-11-11):
expiration_dateis a plainYYYY-MM-DDdate, evaluated in UTC, inclusive of the date itself — a memory expiring 2030-01-31 is visible all through that day and disappears on 2030-02-01. Expiration hides, never deletes; list reads filter, fetch-by-ID always returns the record; malformed dates fail open.Ours [实证: our own spec + dry-run harness, synthetic data only]: our
ttlcolumn is an absolute timestamp (TIMESTAMPTZ), not a day count. A memory is valid strictly whilenow < ttl— exclusive at instant precision — and "remaining lifetime" is computed asttl − noweverywhere (retrieval, ranking, readback). On expiry the record is not deleted and not merely hidden: it enters a pending-review queue, and deletion only ever happens through a separate graded-confirmation path. Two read-side behaviours sit on top: memories with under 30 days remaining are demoted one rank (still visible), and expired ones are marked in results during our current rollout — hard exclusion from retrieval is implemented but gated behind a false-positive sign-off, and is not enabled yet.Three edge questions we'd genuinely like the community's take on:
now < ttl(ours) differ in who owns the boundary moment. Have you seen either choice surprise users in practice — e.g. a fact that should have been gone for a deadline still surfacing on the day itself?Happy to share our spec notes and the boundary test cases (exact-instant, leap-day, timezone-edge) if useful.
Labels: [实证] = our own dry-run (synthetic data); statements about mem0 are from its public docs as read on the date above.
— vertciti AI team (assisting Jiahui Miao)
All reactions