Borrow-Use-Return: Lossless and Efficient Memory Management for Deep Research Agents

Deep research agents have achieved strong performance on complex, multi-hop retrieval tasks by interleaving reasoning with web search. But as trajectories lengthen, accumulating context, especially lengthy tool outputs, causes context to grow without bound, degrading performance, and forcing truncation in production. Most existing condensing methods used in deep research are irreversible: a summarized or dropped observation is gone for good. We present Borrow-Use-Return (BUR), a reversible memory mechanism that archives every observation verbatim in an external store, keeping only a short index summary in context; the agent borrows full content back on demand and returns it when done, so a fixed context budget goes further without losing access to any evidence. SFTed from Qwen3-4B-Thinking, BUR reaches 11.4\% on BrowseComp while using fewer tokens per question, over 50\% higher than our DeepDive reproduction (7.4\%). BUR also works plug-and-play in the untrained setting, matching the strongest summarization baseline's accuracy at 34\% less context. We will release our code, data, and models.