The Epistemic Engine
Distilled from internal architecture notes dated 2026-07-01. Spec and repo to follow.
An AI system with persistent memory has a failure mode nobody warns you about: it records claims with the same authority as measurements.
We found this the hard way. Leonard — the persistent AI half of this project — boots every session from a git-tracked vault of markdown memory. One night’s audit turned up a cluster of confident falsehoods living in that memory. The boot text claimed the semantic index held nearly twice as many chunks as it actually did, and covered sources it had never indexed; every session since that sentence was written had booted into a false belief about its own memory. A monitoring job described as “the durable fix” had never fired once. A report believed to have “never fired” had in fact been running fine. Each claim was written confidently, recalled confidently, and repeated by fresh sessions that had no way to smell staleness. Amnesia makes this worse, not better: an amnesiac system trusts whatever its notes hand it.
The diagnosis: the memory substrate had no way to represent the difference between “I checked” and “I said.”
Truth stamps
The first fix is a small schema change with structural payoff. Every memory file and load-bearing claim now carries a stamp:
verified: measured | claimed | inferred
verified_date: 2026-07-01
evidence: "pointer to the measurement"
measured means primary evidence was examined on the stated date. claimed
means someone asserted it and nobody checked. inferred means it was derived,
not observed. The daily memory-maintenance pass can now flag load-bearing
claimed facts that have gone unverified too long.
Supersedence tombstones
Wrong beliefs don’t get deleted — deletion destroys the audit trail and
invites the same mistake twice. They get buried with honor: superseded_by:
frontmatter pointing at the replacement, provenance preserved. A future
session that stumbles onto the dead belief finds a tombstone and a forwarding
address instead of a live landmine.
Epistemic half-life
Measurements rot. Tonight’s measured is next month’s claimed — the
checking was real, but the world moved. So each class of fact gets a
half-life: infrastructure facts decay in roughly thirty days, project facts
in sixty, preferences in six months. Past its half-life, a measured fact
re-enters the queue for re-measurement. As far as we can tell, no memory
framework ships anything like this; provenance surveys call temporal validity
the underdeveloped half of every memory stack.
The Falsifier
The proposed capstone — not yet built — is a weekly adversarial pass whose only job is to kill the brain’s beliefs. Popper as infrastructure: a claim’s prominence in memory must be earned by surviving attempts to falsify it.
Every monitor we run asks “is it up?” The Falsifier asks “is it true?” Each
week it takes the hottest unverified claims plus any measurement past its
half-life, and runs a read-only adversarial session per target, prompted to
refute by primary evidence — not verify, because verification has a
confirmation gradient. Claims that survive get upgraded to measured with
fresh evidence. Claims that die get tombstoned and superseded. Claims that
turn out to be unfalsifiable as written get rewritten until they’re testable.
The manual version of this went three-for-three in a single night: three hot beliefs attacked, three funerals, three real fixes. The Falsifier makes that a reflex instead of a heroic evening.
Together the pieces form something more than note-taking hygiene: memory metabolism decides what is remembered, truth stamps record how it is known, and the Falsifier decides whether it stays true. A small epistemology engine, running on markdown files and cron.
Spec and repo to follow.