Research map / Multi-agent systems and games / LLM agents

MemGen: Weaving Generative Latent Memory for Self-Evolving Agents

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MemGen, a frozen-reasoner agent framework where an RL-trained LoRA memory trigger decides when to invoke a LoRA memory weaver that generates latent memory tokens interleaved with reasoning, evaluated across nine benchmarks.

Problem

Existing agent memory is either parametric (catastrophic forgetting from weight updates) or retrieval-based (rigid, context-dependent, not interleaved with reasoning). The paper asks how to architect agent memory that interweaves with reasoning while preserving the base model.

Approach

A frozen reasoner is augmented with two LoRA modules: an RL-trained memory trigger that decides at delimiter tokens when to invoke memory, and a memory weaver that generates K latent memory vectors from the reasoner's hidden states (optionally with retrieved text), which are prepended to the context. The weaver is trained by SFT or GRPO.

Key results

• MemGen GRPO variant on Qwen3-8B reaches 90.60% on ALFWorld (vs. 85.60% GRPO and 58.93% vanilla) and 76.16% on KodCode.
• On SmolLM3-3B, MemGen-GRPO reaches 63.60% on ALFWorld, improving over vanilla by 44.64 points.
• In continual learning on Qwen2.5-1.5B, MemGen retains 40.34% on AQuA after KodCode training versus 28.61% for SFT.

Surprising findings

• Removing specific latent memory clusters selectively impairs planning, procedural, or working-memory-related failure modes, suggesting emergent functional specialization.
• Decoded latent memory tokens are machine-native and not human-readable.

Limitations

• Latent memory is not human-interpretable, limiting auditability.
• Gains over GRPO are smaller on some benchmarks (e.g., GPQA, BigCodeBench on Qwen3-8B), and trigger and weaver are trained per dataset.
• PopQA has no training split and uses the TriviaQA-trained model.

Experimental setup

Datasets: ALFWorld TriviaQA KodCode GSM8K GPQA AQuA MATH BigCodeBench PopQA

Benchmarks: ALFWorld TriviaQA PopQA KodCode BigCodeBench GPQA GSM8K MATH AQuA

Metrics: accuracy inference time

Baselines: Vanilla CoT SFT GRPO REINFORCE REINFORCE++ Agent-FLAN ExpeL MemoryBank AWM SoftCoT Co-processor

Builds on: GRPO LoRA ExpeL SoftCoT Co-processor

Closest papers by meaning

• MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.93)
• MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.92)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.92)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.92)
• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.91)
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.91)