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

AdaMEM: Test-Time Adaptive Memory for Language Agents

ICML 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 AdaMEM, a hybrid agent memory that stores successful trajectories offline and synthesizes state-conditioned strategy memory online at each step, plus Step-MFT, a rejection-sampling fine-tuning method that filters strategies by outcome success and action change.

Problem

Existing agent memory systems retrieve static guidance only once at episode start, so agents cannot adapt to intermediate failures or shifting sub-goals in long-horizon tasks.

Approach

Decouple long-term trajectory storage from short-term strategy synthesis: retrieve raw successful trajectories at decision points and generate a fresh natural-language strategy conditioned on the current state, with adaptive refresh modes (low/high effort) and a process-level filter (action change) for fine-tuning the strategy generator.

Key results

• AdaMEM-low reaches 58.2% on ALFWorld unseen vs 52.2% for Synapse and 46.8% for no memory, with WebShop 74.2 vs 71.4 for no memory.
• On HotpotQA with Qwen3-4B, AdaMEM achieves 41.1% success rate vs 40.5% for ReasoningBank and 39.7% for no memory.
• Step-MFT lifts ALFWorld seen success from 54.5% (AdaMEM-high) to 59.8%, while outcome-only filtering degrades WebShop (73.9 vs 76.1).

Surprising findings

• Outcome-only filtering for strategy fine-tuning degrades WebShop performance relative to the training-free AdaMEM-high baseline (73.9 vs 76.1).
• Retrieving raw trajectories without strategy synthesis drops ALFWorld success from 65.5% to 59.3%, and removing long-term memory drops it to 34.8%, below the no-memory baseline.

Limitations

• Strategy inertia: AdaMEM-low relies on self-evaluation and can fail to trigger needed refreshes.
• Long-term memory uses only successful trajectories; failure trajectories are excluded.
• Step-MFT uses rejection sampling SFT rather than online RL, and the process-level signal is a proxy (action change) rather than true step-level utility.

Experimental setup

Datasets: ALFWorld WebShop HotpotQA

Benchmarks: ALFWorld WebShop HotpotQA

Metrics: success rate task score tokens per step inference latency

Baselines: No Memory ReasoningBank Synapse

Builds on: ReAct ReasoningBank Synapse rejection sampling fine-tuning GiGPO

Closest papers by meaning

• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.92)
• MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.92)
• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.92)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.91)
• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.91)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.91)