The $6,000 we couldn't take back

Описание и применение

При удалении компрометированного источника у ИИ‑агента остались все связанные с ним убеждения, планы и действия, что привело к ошибочному платежу. Решение хранит причины каждой памяти в графовой БД FalkorDB и при ретракции удаляет все зависимости, возвращая платеж в список возврата.

Задача

When an AI agent trusts a poisoned source, deleting the source does not remove downstream beliefs, plans, or actions, leading to unintended payments.

Результат

The retraction system removed the contaminated bank account and associated memories, placed the $6,000 payment on a recall list, and blocked a pending $34,500 payment. It prevented further payments that relied on fewer than two independent sources.

Рабочий процесс

• 1. Store each memory with edges to its reasons in FalkorDB.
• 2. When retracting, mark the source’s observations as OUT.
• 3. Mark all downstream memories as OUT tentatively via a variable‑length match.
• 4. Iterate until no memory has all premises OUT, then set surviving memories to IN.
• 5. Preview the retraction on a GRAPH.COPY, commit to apply, or reinstate to revert.

Доказательства

When a source an AI agent trusted turns out to be poisoned, deleting it does almost nothing. Everything the agents built on it stays.

Источник

We store the reasons for every memory as edges in FalkorDB.

Источник

A memory is believed if at least one of its reasons is fully intact.

Источник

In our demo the preview reads: one source, two observations, four beliefs, two plans, three actions. One pending payment ($34,500) is blocked.

Источник