Skip to content

v0.15.8 — Extraction-prompt examples no longer leak names into memory

Choose a tag to compare

View all tags
@hoornet hoornet released this 09 Aug 10:39
· 39 commits to main since this release
v0.15.8
6c3d49e

A one-line prompt fix worth taking if you run a local model.

Fixed

  • Few-shot example names no longer leak into stored memory. The fact-extraction prompt used real names in its identity-category examples ("my name is Jure", "I'm also called Hoornet"). Small local models sometimes copy a few-shot example verbatim into real output, so an Ollama-class extractor could store "User's name is Jure" as an actual fact about you — which then rides into every system prompt via memory injection and resists being talked out of, since nothing you say in conversation outranks something already "remembered". The examples now use generic placeholders.

If you think you've been hit by this, GET /api/memory/{userId} lists what's stored and DELETE /api/memory/{userId}/facts/{factId} removes a single entry. As of 0.16.0 you can also just ask the assistant to forget it.

Reported against the Nives fork (nives#54) — the same prompt text lived here.

Upgrade: no configuration changes required.

Assets 2
Loading