Memory Management for Agents
Methods for managing memory in LLM agents, including consolidation, compression, and learnable memory skills under budget constraints.
53 papers · 4 months
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
June 2026 · University of Science and Technology of China, Microsoft
Memory as state management instead of semantic retrieval improves long-horizon agents
Proposes a memory system based on execution-state tree, outperforming semantic retrieval.
June 2026 · Stanford University, Independent Researcher
First systematic characterization of agent memory systems across long-horizon tasks
Provides the first systematic characterization and taxonomy of agent memory systems.
1 further paper
June 2026 · Johns Hopkins University, Apple
Letting language models decide when to compact their context windows
SelfCompact lets agents dynamically decide when to compact context using summarization and rubrics.
released code
July 2026 · Stanford University
Treating memory management as a trainable cognitive skill boosts LLM performance 2x–4x in long-horizon tasks
Memory management as a trainable skill improves long-horizon performance.
July 2026
Memory consolidation outperforms retention under tight budgets, but retention wins when budgets are loose.
Formal trade-off between memory consolidation and retention under different budget regimes.
Aug 2026
Persistent corpus maps help agents find evidence within tight budgets
Introduces AtlasNav, a persistent corpus map that helps agents locate evidence under budget constraints.
2 further papers
Aug 2026
Query-aware evidence forests improve long-term memory for multimodal agents
Optimizes multimodal memory retrieval as budget-constrained evidence forests.
Aug 2026
Vector memory helps recommendation agents use full interaction histories
Replaces text memory with vector memory for full interaction history usage.
1 further paper
Sept 2026 · Argonne National Laboratory, Columbia University
Coding Agents Need Semantic Measures of Working Memory Performance
Proposes semantic measures and object-aware compression for coding agent working memory.
1 further paper
Sept 2026 · Jilin University
Memory lets traffic-simulation agents build reusable skills without retraining
Uses memory to store reusable skills in traffic-simulation agents, enabling skill reuse without retraining.
released code
3 further papers
Sept 2026
Structured agent memories withstand model upgrades better than compressed notes
Demonstrates that structured knowledge graphs retain memory accuracy better than compressed notes after model upgrades.
Sept 2026
Smarter merging and packing improve LLM memory under tight budgets
Introduces memory merging and packing to improve retrieval under budget, advancing memory management.
3 further papers
Sept 2026 · Beihang University, East China Normal University
Agents can retain old facts without using them indiscriminately
Proposes RD-Forget, a training-free framework separating long-term storage from query-conditioned evidence retrieval.
Sept 2026 · Rutgers University
Kernel-managed memory personalizes multiple AI agents with shorter prompts
Introduces kernel-managed shared memory with controlled retrieval and prompt insertion for multiple agents.
Sept 2026
Refresh schedules, not cache age, determine agent answer staleness
Benchmarks how refresh schedules affect agent answer staleness, informing cache management.
8 further papers
Sept 2026 · Zhejiang University, Alibaba Group
Checkpoint testing reveals hidden weaknesses in self-evolving agent memories
Introduces checkpoint-based evaluation revealing hidden weaknesses in self-evolving agent memories.
Sept 2026 · NLPR&MAIS, Institute of Automation, Chinese Academy of Sciences
Bounded visual workspaces improve multimodal agents’ accuracy and efficiency
Stores visual artifacts in a ledger as an alternative to appending to context, a memory management technique.
5 further papers
Oct 2026 · Apple
Shared scheduler enables realistic training of continual-learning agents
Introduces a shared scheduler that enables realistic memory consolidation and session management for continual-learning agents.
Oct 2026 · Stanford University
Memory Tests Must Separate What Agents Keep From What They Retrieve
Separates retention and selection in memory evaluation, clarifying memory management metrics.
9 further papers
Oct 2026 · University of Tsukuba, Sakana AI
Separate short- and long-term matrix memories improve recurrent network performance
Introduces a recurrent architecture with separate short- and long-term matrix memories updated jointly, advancing memory management for agent systems with explicit temporal segregation.
released code
20 of 53 papers shown