Quote-gate + string checker: agent memory got reliable once the author stopped trusting it
capngreenjeans · reddit · 2026-11-23
A Reddit author describes the memory architecture behind their household-logistics agent on Muse: two layers — a verbatim raw log, plus a curated layer (people/projects/decisions) the agent actually works from.
The key rule is a quote gate: a claim only enters the curated layer if it carries a verbatim quote from the raw log, verified character-by-character by a dumb string-matching script (not a model).
Real test: migrating 63 pages via a cheap Sonnet drafting lane (staging only) cost $12.93 and the checker caught 35 bad quotes — mis-citations or stitched-together real quotes, no invented facts but close enough to have silently corrupted the permanent record.
Superseded claims move to a timestamped section instead of being overwritten; a planted wrong date was caught because the summary disagreed with the log.
Key insight: the pattern only works when the environment hands you the memory layer (your files, your scripts, a checker the agent can't sweet-talk). Anthropic's Duet, which keeps memory behind its own glass, was fundamentally unable to run it. Limits acknowledged: quotes prove existence, not fairness. Partial setup is open source.
More from coding & agent
- User asks Anthropic to publish a full human-agent trace of solving a Millennium Prize problem — _arohan_ · 2026-11-23
- MemTensor launches MemT-1.0-Pro, a 554B memory-native model built with NVIDIA — _akhaliq · 2026-11-23
- Team coordinates AI research agents with a shared research graph — hyperparticle · 2026-11-23
- Kimi K3.1 Pro on Cerebras hits 1375 TPS after 594 agent-run experiments — hyperparticle · 2026-11-23
- Formalizing CFSG at 3x RH scale: Lean is now the bottleneck for AI math — Vjeux · 2026-11-23
- 'Permission is the enemy of autonomy': devs slam Codex approval loops over worktree cleanup — Baconbrix · 2026-11-23