MemPrice: Online Memory Allocation via Adaptive Pricing for Long-Horizon Agents

Key points

• Existing memory mechanisms largely focus on relevance, retrieval, or compression, leaving this allocation trade-off implicit.1
• We introduce MemPrice, which treats active context as a shared resource and dynamically allocates it across memories and representation levels.1
• We show that irreversible compression can force linear worst-case regret against hindsight following its own memory trajectory, and introduce a same-state benchmark to evaluate allocation on MemPrice's realized memory states.1
• We evaluate the framework under matched memory budgets on LongMemEval, LoCoMo, AppWorld, and WebShop across three foundation models, spanning both query-blind memory construction and goal-directed agent context management.1

Sources (1)

• [1]MemPrice: Online Memory Allocation via Adaptive Pricing for Long-Horizon Agents
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 02:11 AM

  Long-horizon LLM agents face a fundamental allocation problem as interaction histories grow: which information to retain, how much context to devote to it, and at what granularity.

  Existing memory mechanisms largely focus on relevance, retrieval, or compression, leaving this allocation trade-off implicit.

Extractive summary: sentences quoted from the sources.