Evidence-backed memory
Most memory layers hand your agent a fact and hope it is still true. MemAPI traces every fact to the exact quote that produced it, gives it a validity window, and retires it the moment it is corrected. Below is a real walkthrough on seeded data - no account needed. Each highlighted span is resolved as a byte-exact range in its source, or flagged when it no longer matches, never reconstructed.
Valid 2026-03-02 → now
Held across both sessions - the follow-up reaffirmed it, so it stays current.
o I can tailor things. User: I prefer dark mode, and please keep answers terse - I skim on mobile. Agent: Noted. Anything about stack? Use
Valid 2026-03-02 → now
A context fact, cited to the exact span of the transcript that produced it.
ile. Agent: Noted. Anything about stack? User: We ship with Next.js 15 and Drizzle ORM on Postgres.
Valid 2026-03-02 → 2026-03-16 · replaced by the correction below
This is the fact a naive accumulate-only store would still serve. Here it LEFT the current set the moment it was corrected - its validity window closed, and retrieval no longer returns it. Its evidence span is shown against the NEWER source revision to demonstrate how a superseded quote is flagged unavailable rather than reconstructed.
The stored quote no longer matches the current source revision content.
Stored quote: I prefer dark mode
Valid 2026-03-16 → now · corrects an earlier value
The correction. It supersedes the dark-mode fact on the same canonical key, with a byte-exact citation to the sentence that stated it.
User: Actually I switched to light mode on the desktop app, dark was straining my eyes. Keep the terse answers though.
This is the same evidence resolution the authenticated dashboard uses - here on public fixture data. Every fact your agents store gets the same citation trail.