A paper making the rounds in AI-agent circles this weekend, "Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents," tackles a failure mode anyone who has run an agent on a task lasting more than a few dozen steps will recognize immediately: as the trajectory grows, decision-relevant facts — what's already been tried and failed, what the environment looks like right now, what the task actually requires — get buried deep in a sprawling context window or pushed out of it entirely, so the agent quietly forgets things it already figured out and repeats mistakes it already diagnosed. The researchers call this "behavioral state decay," and their fix is architecturally simple but a little unusual: rather than trying to make the main agent better at retrieving its own history, they run a second, separate agent alongside it whose only job is memory management. This memory agent watches the unfolding trajectory, maintains a structured memory bank distilled from what's happened so far, and actively decides, turn by turn, whether to inject a memory-grounded reminder into the main agent's context or stay silent and let it proceed. That "decide whether to intervene" framing is the interesting part for builders — most memory systems are passive retrieval-on-request, but this one treats memory as an active intervention the same way a human pair-programmer might interrupt to say "wait, we already tried that." Across two long-horizon benchmarks, adding this proactive memory agent improved task scores by up to 8.3 percentage points without changing the underlying action model at all, meaning it's a wrapper pattern you can layer on top of an existing agent rather than a retraining exercise. For anyone building agents that run long enough to need real memory management — coding agents doing multi-hour refactors, research agents, customer-support agents handling multi-day tickets — the split-agent memory pattern is worth evaluating against whatever ad hoc summarization or vector-retrieval scheme you're currently using, since the core insight of treating memory as an active decision rather than passive lookup generalizes well beyond this specific paper's benchmarks.