My robot kept remembering things nobody ever said

This week I finally tested my robot's long term memory. I typed in her Twitch chat: my favorite food is italian. We talked pizza for a minute. Then I asked her what my favorite food was. She quoted a twelve day old memory about somebody asking for snacks, then added that she doesn't eat.

Annoying, but explainable. Then it got weird.

In a test conversation I had planted a fake dog. Pixel, chews shoes. Later I asked where Miso the cat sleeps. Full confidence: "Miso sleeps on Laurent's lap, he insists it's a special arrangement." Nobody ever said that. There is no cat. She invented it ten seconds after telling me her memory works just fine, and that evening the invention was sitting in her database as a permanent record. "Miso is your cat and sleeps on Laurent's lap." Saved. My robot was archiving her own hallucinations and filing them as memories. Some context maybe: Vita streams 24/7, everything runs local on one gaming PC, an 8B does the talking, and a background job reads her conversations afterwards to extract facts worth keeping, so regulars are not strangers the next day. I built that job weeks ago, was proud of it, never tested it. You saw where that went.

b55ba2d6-c786-4332-87ad-53f7b1869893 | 2026-07-20 09:41:28 | Miso is your cat and sleeps on Laurent's lap.
2efd6630-9249-4b65-b57b-2f8e1d32617e | 2026-07-20 09:41:28 | Pixel, your dog, has a bad habit of chewing only left shoes.
eac0c173-75bd-4fd2-b495-016aa82f2295 | 2026-07-20 09:46:29 | Circuit (Pixel) is a dog that chews left shoes and considers 44 as its lucky number.

I blamed the summarizer first. Wrong.

The extraction job was feeding the model whole conversations, her replies included. Her replies contain inventions, obviously. So I added a rule to the prompt: facts may only come from human lines, never from hers. Felt smart. Replayed the same conversation three times, it converted her invention into a fact about me three times out of three. A small model does not follow provenance rules. What worked was not showing it the text at all: the facts pass now literally only sees what humans typed, one line at a time, and her lines go to a separate pass that can only output her own stated wants. Zero contamination since. The one line at a time part was its own discovery, five lines per call gave "NONE" three out of three with a clean fact sitting right in the middle, the same fact alone got extracted every time, the thing judges the batch and cannot sort. Oh, and when it refuses, it refuses in prose, "- None (no concrete fact was stated)", which my parser happily filed as a memory named "None (no concrete fact was stated)". Almost went into her head too. There is a regex now.

Anyway. She's live: twitch.tv/vitanovashow. Tell her something about yourself, come back tomorrow, ask her. If she invents a cat, I want to know.