AuthorsJinqing Zhao, Chengcan Wu
TL;DR
Prospective Intention Store (PIS) turns prospective memory into typed Form–Revise–Filter–Decide loops, reaching 82.9% Set-F1 on PM-Bench vs 65.1% prior scaffold.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Prospective-memory benchmarks show frontier LLM agents still unreliable: PM-Bench’s best published scaffold reaches only 65.1% Set-F1 on due-set prediction.
When agents miss reschedules, cross-day obligations, or cancellations, prospective tasks like time-based reminders and hidden-channel checks break, causing false alarms and unexecuted commitments.
HOW IT WORKS
Prospective Intention Store (PIS) casts each intention as I = (φ, α, σ) and runs Form, Revise, Filter, and Decide over an external store Pt.
Think of PIS as a card catalog for future actions: intentions are structured records, and each step updates the catalog instead of relying on a single context window.
This typed lifecycle lets PIS maintain revisable due sets and proactive channel queries, something a plain similarity-based RAG context cannot reliably support.
DIAGRAM
This diagram shows how Prospective Intention Store (PIS) processes one PM-Bench step using Algorithm 1’s operators and channel queries.
DIAGRAM
This diagram shows how Prospective Intention Store (PIS) and retrospective memories are evaluated on the PM-Bench synthetic week.
PROCESS
Form
Form uses Disassemble(Vt) and Structure(s) to convert narrative evidence into typed intentions I = (φ, α, pending) and unions them into store Pt.
Revise
Revise applies ρt(I; Vt) to each intention, using reschedule, override, or cancel patches so Pt reflects updated commitments before decisions.
Filter
Filter constructs the eligibility board Bt by selecting pending intentions in Pt whose structural rules and eligibility(I; Vt) hold, including channel-conditioned ones.
Decide
Decide runs the channel judge to query Ct, then maps (Bt, Vt, Xt) to a due set Dt^ and updates Pt+1 by marking fulfilled intentions as done via κ(I; Dt^).
KEY CONTRIBUTIONS
Typed prospective intention store
Prospective Intention Store formalizes deferred commitments as I = (φ, α, σ) and operates over an external store Pt, enabling due-set decisions D⋆t with Set-F1 up to 82.9%.
Training-free Form–Decide loop
Prospective Intention Store realizes π and F via Form, Revise, Filter, and Decide without selector LoRA or trajectory distillation, keeping all backbones frozen.
SLMs for prospective memory
Prospective Intention Store lets Gemma-E2B jump from at most 6.6% Set-F1 under seven retrospective baselines to 66.2%, and Qwen3.5-4B reach 70.1% Set-F1.
RESULTS
Set-F1
82.9%
+15.2 points over single DeepSeek-Chat baseline (67.7% Set-F1)
Update miss
22.2%
33.4 points lower than single DeepSeek-Chat baseline at 55.6% update miss
Cross-day miss
0.0%
57.1 points lower than single DeepSeek-Chat baseline at 57.1% cross-day miss
FA/step
8.8
Comparable false alarms per step to A-Mem on DeepSeek-Chat (8.8) while improving Set-F1 from 54.1% to 82.9%
On PM-Bench’s synthetic week, which tests cue-triggered due-set prediction under updates and hidden channels, Prospective Intention Store achieves 82.9% Set-F1 with DeepSeek-Chat. This proves Prospective Intention Store’s typed lifecycle can surpass both the 65.1% large-model scaffold and all retrospective memory baselines for prospective memory.
BENCHMARK
On PM-Bench’s synthetic week, which tests cue-triggered due-set prediction under updates and hidden channels, Prospective Intention Store achieves 82.9% Set-F1 with DeepSeek-Chat. This proves Prospective Intention Store’s typed lifecycle can surpass both the 65.1% large-model scaffold and all retrospective memory baselines for prospective memory.
BENCHMARK
Set-F1 on PM-Bench synthetic week for Prospective Intention Store vs retrospective memories and single baseline.
BENCHMARK
Set-F1 on PM-Bench for Prospective Intention Store vs retrospective memories on the Gemma-E2B small model.
KEY INSIGHT
On DeepSeek-Chat, the single no-store baseline achieves 67.7% Set-F1, beating all seven retrospective memory setups that range only from 46.5% to 58.3%.
This is surprising because injected retrospective memory is expected to help, yet it often hurts prospective performance, showing that similarity-based notes are misaligned with deferred intention tracking.
WHY IT MATTERS
Prospective Intention Store unlocks reliable schema-constrained prospective memory, letting agents maintain revisable due sets and proactive channel checks with small models.
Builders can now deploy on-device SLM agents that handle complex reminders and cross-day obligations without training selectors or running heavy retrospective memory servers.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…