Agent memory design: entity resolution, dated facts, and trust-ranked truth
jasonkneen · x · 2026-10-10
In a quoted thread, @omarships breaks down the data layer behind their agent's "brain":
- Entity resolution: one record per person, company, deal, and invoice; matching on hard IDs first (CRM/Stripe IDs, emails, domains), names second—two John Smiths with different emails stay separate.
- Facts are dated: each fact stores when it became true and when a source said it; newest wins, then most-trusted source. A proposal never overrides a fact, and your own correction beats lower-trust claims.
- Freshness & lifecycle: stale facts get re-read live before the agent acts on them; deleting a source deletes the fact; inbox-private facts stay visible only to you.
- Motivation: plain search ranked outdated values first 3 times when old and new coexisted.
jasonkneen shared it with a joke that it's a simulation with parts sped up.
More from AGI Musings
- The teamsters lesson: automation adjustment is least painful when newcomers just stop entering — soumitrashukla9 · 2026-10-23
- Researcher pushes back on AI cyber fear: 'labs selling defense by spreading fear' — gerardsans · 2026-10-23
- Nina Schick: compute shortage is itself a tremendous cybersecurity risk — NinaDSchick · 2026-10-23
- Investor: If AI is a bubble, it's the weirdest one ever seen — robleclerc · 2026-10-23
- Chamath: Banks will block AI agents via terms of service before anything else — nicolechirps · 2026-10-23
- AI researcher: Claude finally finished tedious research tasks I gave up on years ago — MikePFrank · 2026-10-23