Agentic Memory: Agents Begin to Unify the Writing, Retrieval, and Forgetting of Long- and Short-Term Memory
The Agentic Memory paper submitted on January 5, 2026 proposes AgeMem, which unifies long- and short-term memory operations into learnable tool actions for agents, and reports consistent improvements over strong memory baselines on five long-horizon benchmarks.
Agent memory has long relied on fixed heuristics for summarization, vector retrieval, and external controllers. AgeMem incorporates storage, retrieval, update, compression, and deletion into the same policy, enabling memory management to be jointly optimized with task feedback for the first time.
AgeMem represents long- and short-term memory operations as callable actions, and uses three-stage progressive reinforcement learning with stepwise GRPO to mitigate the sparse, non-continuous rewards from memory operations. The paper compares task success rate, memory quality, and context efficiency across multiple foundation models and five types of long-horizon tasks.
The memory layer is evolving from a peripheral plugin in agent frameworks to a part of the trainable policy. Future platform differentiation will depend not only on the storage engine but also on the quality of policies for when to write, retain, compress, and forget.
When procuring or building long-horizon task agents, evaluate completion rate, memory contamination rate, context cost, and deletion capability simultaneously, avoiding decisions based solely on the feature label 'supports long-term memory'.
Further validation is needed for stability across models, tools, and real enterprise tasks, as well as measurement of reinforcement learning costs, error memory accumulation, privacy deletion, and auditability.