Making Prospective Memory SLM-Shaped: Typed Intention Stores for Small-Model Agents
Jinqing Zhao and Chengcan Wu argue that prospective memory, carrying out a deferred intention at the right future cue, is schema-constrained state tracking rather than open-ended reasoning, and show that typing the action space lets small models beat the published large-model scaffold.
Ask this paper
The reframe is the contribution. If the loop is state tracking, lifecycle logic belongs in code and only scoped language work belongs to the model. The Prospective Intention Store makes that split explicit.
Training-free. No selector fine-tuning and no trajectory distillation. The scaffold is agentic and works off the shelf, which lowers the bar for adoption considerably.
The small-model result is the striking one. On PM-Bench, Gemma-E2B reaches 6.6% Set-F1 at best under seven retrospective memory methods, but 66.2% with PIS. DeepSeek-Chat with PIS reaches 82.9% against the best published scaffold's 65.1%.
Retrospective memory is the wrong tool. PIS reaches 70.1% Set-F1 where retrospective memory methods stay at 54.4% or below, which supports the claim that remembering the past does not substitute for scheduling the future.
Abstract
Prospective memory means carrying out a deferred intention at the right future cue while other work continues. Benchmarks now isolate it as an agent skill, yet frontier LLMs still struggle: the best published PM-Bench scaffold reaches only 65.1% Set-F1. We argue that this loop is schema-constrained state tracking rather than open-ended reasoning, and that small models can execute it when the action space is typed. We propose the Prospective Intention Store (PIS) that puts lifecycle logic in code and scoped language work on the model. The scaffold is agentic and training-free: no selector fine-tuning and no trajectory distillation. On PM-Bench, DeepSeek-Chat with PIS reaches 82.9% Set-F1. On Gemma-E2B, Set-F1 is only 4.2% without a store and at most 6.6% under seven retrospective memories, while PIS reaches 66.2%. PIS further reaches 70.1% Set-F1, where retrospective memory methods stay at most 54.4%. PIS sets a new state of the art on this benchmark and enables small models to surpass the published large-model scaffold.