Towards mental time travel: a hierarchical memory for reinforcement learning agents

Reinforcement learning agents often forget details of the past, especially\nafter delays or distractor tasks. Agents with common memory architectures\nstruggle to recall and integrate across multiple timesteps of a past event, or\neven to recall the details of a single timestep that is followed by distractor\ntasks. To address these limitations, we propose a Hierarchical Chunk Attention\nMemory (HCAM), which helps agents to remember the past in detail. HCAM stores\nmemories by dividing the past into chunks, and recalls by first performing\nhigh-level attention over coarse summaries of the chunks, and then performing\ndetailed attention within only the most relevant chunks. An agent with HCAM can\ntherefore "mentally time-travel" -- remember past events in detail without\nattending to all intervening events. We show that agents with HCAM\nsubstantially outperform agents with other memory architectures at tasks\nrequiring long-term recall, retention, or reasoning over memory. These include\nrecalling where an object is hidden in a 3D environment, rapidly learning to\nnavigate efficiently in a new neighborhood, and rapidly learning and retaining\nnew object names. Agents with HCAM can extrapolate to task sequences much\nlonger than they were trained on, and can even generalize zero-shot from a\nmeta-learning setting to maintaining knowledge across episodes. HCAM improves\nagent sample efficiency, generalization, and generality (by solving tasks that\npreviously required specialized architectures). Our work is a step towards\nagents that can learn, interact, and adapt in complex and temporally-extended\nenvironments.\n

Paper

Similar papers

© 2026 NYSGPT2525 LLC