AuthorsYefan Zhou, Yang Li, Zeyu Leo Liu et al.
TL;DR
Just-in-Time Memory (JITMEM) defers memory curation to read time, letting a GRPO-trained curator synthesize task-adaptive payloads that beat SkillOS by up to 16.3 SR points on WebShop.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing agentic memory systems curate at write time, producing query-independent summaries that must serve many future tasks. This prematurely and irreversibly discards trajectory details needed later.
This breaks long-horizon LLM agents, forcing them to guess what matters before seeing downstream tasks and creating a difficult credit-assignment problem over many future interactions.
HOW IT WORKS
JITMEM’s core mechanism combines a Memory Bank, Retriever, Memory Curator, and frozen Agent Executor so only raw trajectories are stored and curated at read time. The UPDATE gate uses the executor-as-judge to keep only successful trajectories.
You can think of JITMEM like a card catalog plus librarian: the Memory Bank keeps full books, the Retriever finds shelves, and the Memory Curator summarizes exactly the chapters relevant to the current reader’s question.
This read-time design lets JITMEM tailor payloads to each task, avoid premature information loss, and learn curation directly from immediate task rewards in a way a plain context window cannot.
DIAGRAM
This diagram shows how JITMEM retrieves raw trajectories and synthesizes a task-adaptive payload at inference time before executing the current task.
DIAGRAM
This diagram shows how JITMEM trains the Memory Curator with GRPO using a fixed training bank and a frozen Agent Executor.
PROCESS
Retrieve
At each task xt, JITMEM uses the Retriever R over the Memory Bank Mt to fetch the top k raw trajectories most relevant to the task description.
Curate
The Memory Curator πϕ reads xt and the retrieved trajectories, then synthesizes a compact task adaptive payload pt that extracts strategies and guidance tailored to xt.
Execute
The frozen Agent Executor πL receives xt and pt in its prompt, uses the curated payload as guidance, interacts with the environment, and produces a trajectory ξt and task reward rt.
Update
JITMEM runs an executor-as-judge to assess correctness and uses UPDATE to append ξt to the Memory Bank only if the trajectory passes the quality gate.
KEY CONTRIBUTIONS
Read-time curation enables task-adaptive memory
JITMEM defers curation to read time so the Memory Curator sees the current task xt and can distill different payloads from the same raw trajectory for different tasks.
Immediate credit assignment via GRPO
JITMEM trains the Memory Curator with GRPO, using the same task’s reward rt to update curation without task grouping, collapsing credit assignment to a single interaction.
Empirical gains over write-time memory
On ALFWorld, WebShop, and τ2-bench, JITMEM improves success rate by up to +16.2 points over SkillOS and reduces executor steps by 28.4%–31.4% compared to write-time methods.
RESULTS
ALFWorld SR
77.4%
+16.2 over SkillOS with Qwen3-8B executor (61.2%)
WebShop SR
32.8%
+16.3 over SkillOS with Qwen3-8B executor (16.5%)
WebShop SR Gemini
50.5%
+9.2 over SkillOS with Gemini-2.5-Pro executor (41.3%)
τ2-bench Macro SR
73.4%
+3.0 over ReasoningBank with GPT-5.4 curator (70.4%)
These results come from ALFWorld, WebShop, and τ2-bench, which test embodied control, web shopping, and multi-domain tool-use. The gains show JITMEM’s read-time curation improves task success across diverse agent benchmarks compared to ReasoningBank, MemP, and SkillOS.
BENCHMARK
These results come from ALFWorld, WebShop, and τ2-bench, which test embodied control, web shopping, and multi-domain tool-use. The gains show JITMEM’s read-time curation improves task success across diverse agent benchmarks compared to ReasoningBank, MemP, and SkillOS.
BENCHMARK
Success rate on ALFWorld comparing JITMEM to write-time memory baselines using Qwen3-8B as executor.
BENCHMARK
Success rate on WebShop comparing JITMEM to write-time memory baselines using Qwen3-8B as executor.
KEY INSIGHT
Even an untrained JITMEM curator is competitive with or surpasses RL-trained write-time methods, reaching 61.0% SR on WebShop with Gemini-2.5-Pro versus 41.0% for SkillOS.
This is surprising because SkillOS uses a stronger curator model and RL training, yet JITMEM’s simple task-adaptive read-time curation alone yields a +20.0 point success-rate advantage.
WHY IT MATTERS
JITMEM unlocks reconstructive, task-conditioned episodic memory for LLM agents, letting them reuse full past trajectories without committing to fixed write-time summaries.
Builders can now plug a trained Memory Curator into different executors and get immediate, task-specific guidance from streaming experience, without designing task groups or complex write-time policies.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Xiaoyang Li, Yiqi Wang et al.
arXiv 2026 · 2026
Correlated Promotion Benchmark (CPB) combines CPB-Static, CPB-Live, a gold admission rule, lineage collapse, and a governance rule to stress-test epistemic admission in shared agent memory. On CPB-Live, the governance rule keeps damage shares between 0.112 and 0.152 and false adoption between 0.06 and 0.09, while majority vote and LLM judges often match share-all’s false adoption.
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.
Answers use this explainer on Memory Papers.
Checking…