Typed intention stores let an on-device model beat the best published prospective-memory scaffold

arXiv·medium signalXBlueskyLinkedInCopy link

Prospective memory, carrying out a deferred intention at the right future cue while other work continues, is schema-constrained state tracking rather than open-ended reasoning, so the Prospective Intention Store puts the lifecycle logic in code and leaves only scoped language work to the model. On PM-Bench, where the best published scaffold reached 65.1% Set-F1, DeepSeek-Chat with PIS reaches 82.9%. The result that matters for local deployment is Gemma-E2B: 4.2% Set-F1 with no store, at most 6.6% with seven retrospective memories, and 66.2% with PIS. The scaffold is training-free with no selector fine-tuning and no trajectory distillation.