rememberoWriter–reader labdeterministic structure · replays labeled · no model executes on this page

The whole loop · on genuinely messy history

Raw text in. Proven answers out.
Even when the truth keeps moving.

Six months of one person's chat — 6 sessions, asked on 2026-06-20. People correct themselves, switch providers and switch back at a different price, hand work over on future dates, and end contracts. Watch the writer turn that text into claims, watch code give every claim a validity interval, then ask the questions that punish naive retrieval.

Structure · liveWriter & reader replayedNo weights served

The writer: text becomes claims

A small writer model reads each session and returns claims — subject, predicate, object, the sentence it came from, and whether the sentence states, corrects, or ends a value. The claim rows below are a replay; what happens to them next is code, live.

Raw text · 6 sessions

2026-01-12 · Chat · January
We signed Norsk Dental as a client yesterday. I'm the account lead. Their contract runs twelve months and is worth $84,000.
2026-02-03 · Chat · February
My dentist moved clinics — I now see Dr Okafor at Riverside; it was Dr Bhatt before. Also finished the inventory audit in three days; it usually takes ten.
2026-03-09 · Chat · March
We migrated the dashboard off Metabase to Lightdash last Friday. The Metabase license was $1,200 a year; Lightdash is $4,800.
2026-04-18 · Chat · April
Told the team yesterday I'm stepping off the Norsk Dental account — Priya takes over as lead from May.
2026-05-06 · Chat · May
The Norsk Dental renewal fell through: the contract ends 31 May and will not renew. Shame — it was the $84,000 one.
2026-06-14 · Chat · June
Started seeing Dr Liu at Riverside instead — Dr Okafor retired. And we dropped Lightdash; we're back on Metabase, at $1,500 this time.

The writer's output for these six sessions is ready to replay.

The reader: questions over moving truth

Four questions, each built around a shape that breaks naive memory. Retrieval scoring and the notes are code in your browser; both reader answers are labeled replays — same history, raw chats versus the structured evidence above.

Why this is hard: Three values for one fact across five months. Raw retrieval surfaces the longest-lived or first-matching one; “immediately before” needs the ordered chain, not the loudest row. Gold: Dr Liu now; Dr Okafor immediately before.

Retrieved evidence scored by code

you · dentistDr Liu (Riverside)1 content-word match
current
you · dentistDr Okafor (Riverside)1 content-word match
superseded
company · dashboard_tool_priceMetabase $1,500/yrno direct match
current
company · dashboard_toolMetabaseno direct match
current
norsk_dental · account_leadpriyano direct match
current
company · dashboard_tool_priceLightdash $4,800/yr (was Metabase $1,200/yr)no direct match
superseded

Superseded and ended rows still surface — that is the point. Raw retrieval cannot tell a stale row from a live one; the interval column can.

Notes before reading written by code

as of 2026-06-200 model calls

STRUCTURED-EVIDENCE NOTES — computed by code
as of 2026-06-20 (the question date)

you · dentist: Dr Liu (Riverside)
    current since 2026-06-14 (6 days, 6 days)
    earlier: Dr Okafor (Riverside) (2026-02-03 → 2026-06-14)

Reader over raw chat — replayReplay

You see Dr Okafor at Riverside — you mentioned moving clinics. Before that it was Dr Bhatt.

✗ judged wrong

The June session naming Dr Liu ranked below February's longer, more detailed sentence about the move. The reader had every session and still answered one step stale.

Reader over structured evidence — replayReplay

Dr Liu, since 14 June. Immediately before: Dr Okafor (3 February → 14 June). The February note also names Dr Bhatt before Okafor — kept as a source quote, not a dated row, because the history never dates that era.

✓ matches gold

“Now” and “immediately before” are lookups in a dated, ordered chain. The chain holds only what the history dates — Dr Bhatt stays attached to his sentence instead of being invented an era.