Skip to content
memory with a receipt

The account memory that shows its work.

Two years of an Acme Org rollout — 45 documents, 2,400-odd turns and messages. Receipt reads them into a claim memory where every proposition carries its sources, distinguishes what was proposed from what was agreed, knows what has been overtaken, forgets a person on request, and reviews the account unprompted.

45

documents

2426

live segments

106

memory claims

272

bound citations

12

signals on review

Ask the account →

Answers composed from the memory with claim-level receipts: document, position, speaker, date — and an explicit “the archive does not establish this” when that is the truth.

Review the account 12 signals

The agent’s own review: obligations with no completion evidence, a plan built on an agreement nobody made, wrong-when-written records, and the unverified repetition behind the board number.

Read the commitment ledger →

Who proposed, who agreed, what happened after — with the trail that proves each step.

Erase a person →

Real deletion across the working corpus, citations, claims, commitment trails, summaries, caches and indexes — with a receipt. Other people stay named.

Latest signals

all 12 →
red since 2026-06-26

The warehouse data field has been open since April 2024 — and the chain-wide case now depends on it

red since 2026-06-16

Three weekly “no errors” reports are evidence of job completion only — and the data does not reconcile

red since 2026-03-28

The bake-off split by store format was requested in January and was still only in Sofia’s spreadsheet in March

Receipt — built for the RELEX track at AaltoAI Hackathon 2026. Storage and inference entirely on this host (Helsinki, EU); the archive lives in a local SQLite store; no external API is called at any point.