AuthorsDongge Han, Camille Couturier, Daniel Madrigal Diaz et al.
TL;DR
LEGOMem uses modular full task memories and subtask memories to guide multi agent planning and tool use, boosting OfficeBench success by up to 13.38 percentage points over no memory.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Current multi agent LLM systems are largely stateless and transactional, solving each task from scratch without reusing prior experience.
This limits procedural memory for workflow automation, so orchestrators and task agents repeatedly fail complex OfficeBench tasks instead of improving execution skills over time.
HOW IT WORKS
LEGOMem introduces procedural memory bank, full task memories, and subtask memories, plus the LEGOMem-Dynamic and LEGOMem-QueryRewrite variants for fine grained retrieval.
You can think of LEGOMem like a card catalog of past workflows, where orchestrators pull whole task cards and agents pull Lego like subtask bricks.
This design lets LEGOMem reuse structured execution traces beyond a plain context window, enabling better planning, delegation, and tool use in multi agent workflows.
DIAGRAM
This diagram shows how LEGOMem retrieves memories and allocates them to orchestrator and task agents during Algorithm 1 multi agent execution.
DIAGRAM
This diagram shows how LEGOMem is trained and evaluated on OfficeBench, including memory curation, team configurations, and ablation settings.
PROCESS
Offline memory construction
LEGOMem runs the LLM team without memory on 148 training tasks, extracting 93 successful trajectories into the procedural memory bank and 250 subtask memories.
Memory retrieval and allocation
For a new task, LEGOMem embeds the task description, retrieves top K full task memories, and assigns corresponding subtask memories to each task agent.
Multi agent execution loop
LEGOMem orchestrator generates an initial plan, selects task agents, and uses subtask memories to guide tool use and summarize observations back to the orchestrator.
Replanning and final response
If progress stalls, LEGOMem orchestrator replans using full task memories, then continues until the environment state satisfies the goal and returns the final answer.
KEY CONTRIBUTIONS
LEGOMem modular procedural memory
LEGOMem introduces a procedural memory bank with full task memories and subtask memories, enabling role aware reuse of 93 successful trajectories and 250 subtask traces.
LEGOMem variants for retrieval
LEGOMem defines LEGOMem-Dynamic and LEGOMem-QueryRewrite to explore subtask level retrieval, including just in time retrieval and query rewriting for agent specific memory banks.
Systematic study of memory placement
LEGOMem provides ablations on orchestrator memory and task agent memory, showing orchestrator memory raises LLM team success from 45.83% to 53.29% even without agent memories.
RESULTS
Overall success rate
58.44%
+12.61 over No memory (LLM team overall 45.83%)
SLM overall success
38.16%
+13.38 over No memory SLM team (24.78%)
Hybrid overall success
50.22%
+14.91 over No memory Hybrid team (35.31%)
Level 3 LLM success
47.46%
+14.13 over No memory LLM Level 3 (33.33%)
These metrics come from OfficeBench, a 300 task office automation benchmark split into 148 training and 152 test tasks, evaluating multi step workflows across three difficulty levels. The gains show that LEGOMem makes multi agent systems substantially more reliable than stateless baselines by reusing procedural memories.
BENCHMARK
These metrics come from OfficeBench, a 300 task office automation benchmark split into 148 training and 152 test tasks, evaluating multi step workflows across three difficulty levels. The gains show that LEGOMem makes multi agent systems substantially more reliable than stateless baselines by reusing procedural memories.
BENCHMARK
Overall success rate on OfficeBench test set for SLM team configurations.
BENCHMARK
Overall success rate on OfficeBench for Hybrid (LLM + SLM) team under different memory placements.
KEY INSIGHT
Hybrid teams with LEGOMem-QueryRewrite reach 50.22% overall success, surpassing the memory less LLM team at 45.83% on OfficeBench.
This is surprising because it breaks the assumption that larger LLM only teams always dominate, showing procedural memory can compensate for smaller task agents.
WHY IT MATTERS
LEGOMem lets multi agent systems accumulate reusable procedural skills, so orchestrators and task agents improve planning, delegation, and tool use over time.
Builders can now deploy smaller or hybrid teams that match or exceed larger LLM teams by plugging in LEGOMem as a modular procedural memory layer.
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…