简体中文 · English

LEGOMem: Composable procedural memory

Fidelity: Implements procedural units for orchestrator and execution layers, extraction from successful trajectories, and cross-episode reuse. The current PlanBench mini-suite is not the paper's multi-agent OfficeBench environment.

Paper information

Field	Details
Paper	LEGOMem (arXiv 2510.04851)
Institution	Microsoft Research
First public release	2025-10-06
Author code	No public repository found on the paper page as of 2026-07-27.
Local adapter or method key	legomem
Local reproduction code	src/auto_research/agent_research/

Original paper summary

Background and main changes

Whole successful trajectories are hard to transfer to new tasks, while one global memory mixes task decomposition with tool execution. LEGOMem breaks experience into composable procedural units. Orchestrator memory stores decomposition and delegation, agent memory stores action templates, and the units are recombined at runtime for each new task.

flowchart LR
    T["New task"] --> O["Orchestrator memory"]
    O --> D["Decomposition + delegation"]
    D --> A1["Agent procedure unit"]
    D --> A2["Agent procedure unit"]
    A1 --> V["Verified plan"]
    A2 --> V
    V --> W["Write back successful units"]
    W --> O

Key figure from the paper

Figure 3: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations

Successful trajectories \(\tau\) are split into orchestration and execution units, retrieved for the task, and combined:

\[ \mathcal{M}=\mathcal{M}_{\mathrm{orch}}\cup\mathcal{M}_{\mathrm{agent}}, \qquad \pi(T)=\operatorname{Compose}\!\left( \operatorname{Retrieve}(T,\mathcal{M}_{\mathrm{orch}}), \operatorname{Retrieve}(T,\mathcal{M}_{\mathrm{agent}}) \right). \]

Only task-verified procedural units enter long-term memory.

Offline and online results

The paper analyzes complementary effects of orchestrator and agent memory on OfficeBench. Orchestrator memory primarily improves decomposition and delegation, while agent memory improves execution. There is no production online A/B test, and the public abstract does not give a single figure suited to citation out of the results table.

Local reproduction

planbench-mini uses verifiable structured plans. The implementation extracts action/domain units from successful plans, reuses high-level decomposition and fine-grained actions separately, and records reused_plans.

auto-research agent-eval --method legomem \
  --benchmark planbench-mini --episodes 120 --seed 42

Metric	LEGOMem
Joint success	1.0000
Average cost	1.1200
Reused plans	108
Final memory size	12

Stable metrics: agent-mini-suites-seed42.json.

Scope

The local run has no real Office applications, LLM multi-agent delegation, or error recovery. It demonstrates extraction, verification, and cross-episode reuse of procedural units, not reproduction of the OfficeBench results.