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

Learning Query-Aware Budget-Tier Routing for Runtime Agent Memory

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 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 BudgetMem, a runtime agent memory framework where each memory module has Low/Mid/High budget tiers and a PPO-trained lightweight router selects tiers per module to trade off F1/Judge performance against token cost on LoCoMo, LongMemEval, and HotpotQA.

Problem

Runtime agent memory extraction is costly and offers little explicit control over the performance-cost trade-off, while offline query-agnostic memory construction can discard query-critical information.

Approach

Expose each memory-pipeline module at three budget tiers (implementation, reasoning, or capacity tiering) and learn a shared query-aware router with RL that picks a tier per module under a cost-aware reward with variance-based reward-scale alignment.

Key results

• Under the performance-first setting with LLaMA-3.3-70B, BudgetMem-Cap reaches 60.50 Judge on LongMemEval versus 48.51 for the strongest baseline LightMem.
• Average over the three datasets with LLaMA-3.3-70B: BudgetMem-Cap 45.72 F1 / 59.99 Judge vs LightMem 35.45 F1 / 49.21 Judge.
• On HotpotQA with Qwen3-Next-80B, BudgetMem-Rea reaches 70.83 Judge at cost 0.17, versus BudgetMem-Cap 72.08 Judge at 0.22.

Surprising findings

• Without reward-scale alignment (lambda=0.3), the router collapses to the Low tier for most modules, sharply degrading answer quality.

Limitations

• Evaluated only on dialogue/QA memory tasks with a fixed module set (filter, entity, temporal, topic, summary).
• Cost is measured in API token pricing and the LoCoMo cost differences are small because histories are short.
• Reasoning tiering has narrow cost spread, offering less range for low-budget or high-budget extrapolation.

Experimental setup

Datasets: LoCoMo LongMemEval HotpotQA

Benchmarks: LoCoMo LongMemEval HotpotQA

Metrics: F1 LLM-as-a-judge accuracy monetary cost (token-based) latency

Baselines: ReadAgent MemoryBank A-MEM LangMem Mem0 MemoryOS LightMem

Builds on: PPO LightMem Contriever

Closest papers by meaning

• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.90)
• Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents (NeurIPS 2026 · similarity 0.90)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.90)
• AdaMEM: Test-Time Adaptive Memory for Language Agents (ICML 2026 · similarity 0.90)
• MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents (ICLR 2026 · similarity 0.89)
• Mem-T: Densifying Rewards for Long-Horizon Memory Agents (ICML 2026 · similarity 0.89)