Memory you can live by.
Maximus runs on a longitudinal memory architecture built to support active schedules in daily life — is it accurate enough to reduce user cognitive load? — while also prompting and supporting human change over a long arc of time.
To do this, I broke down memory and prompting into four layers, arranged it across six key life surfaces, and made the record durable.
The four-layer mental model
Every failure in the memory pipeline routes to exactly one of four layers. Each fails differently and is fixed differently — so a call that feels wrong gets a procedure, not a mystery.
1 · Capture
Voice becomes transcript. Its failure looks like garbled speech-to-text or dropped audio — and the fix lives in the capture stack, never downstream.
2 · Storage
The transcript becomes atomic facts — goals, money, commitments, one row each. Its failure looks like a stated fact never recorded, or recorded with an invented time. The fix is in extraction.
3 · Synthesis
The facts become the one-page snapshot the next session opens with. Its failure looks like a commitment that exists in the record but never made the page. The fix is in synthesis, not storage.
4 · Generation
The snapshot becomes the conversation. Its failure looks like the right context said badly, or at the wrong moment. The fix is in the prompt, not upstream.
The engine
- Append-only. A fact is never edited and never deleted. A correction retires the old version with a recorded reason and writes a new one on the same chain — the audit trail is the schema itself.
- Provenance on every fact. Which call, which turn, which model, which prompt version wrote it. "Why does it think this?" is a query, not an investigation.
- Honest abstention. When the record can't know — "three o'clock" with no AM or PM — it stores the fact undated and asks. A wrong answer delivered confidently is the one unrecoverable failure for a system of record.
The loop
A call becomes a transcript; extraction turns the transcript into individual facts; a matcher decides new-thing-or-update; the current facts are synthesized into a one-page picture of the user's life; the next call — and the next morning's text — open already knowing it. Reminders ride on the facts themselves, so a correction propagates for free. Every turn of every call logs latency, cost (in billionths of a dollar), and errors.
The governance is the product
Every line of code was typed by an AI agent. What makes that shippable is the institution around it: every pull request externally reviewed before merge; every bug fixed gets a permanent regression test; every new seam mutation-tested — broken on purpose, to prove a test notices; language changes graded against real data before they ship; every decision recorded, dated, in an append-only log.
Maximus is armed automation acting unprompted over real people's lives. It sends texts, fires reminders, retires records. The discipline isn't overhead. It's the product.
The record
Counted from the repositories, not estimated.
The same engine runs two more builds: Dispositive, an evidence engine for private-equity diligence, and First Person, a measurement instrument for medication response.
The writing carries the long version — the architecture, the instruments, the stumbles.