A new architecture splits agentic memory into two jobs: one agent acts, a second agent decides when to interrupt it. The split, described in a paper posted to arXiv on July 9, targets a failure mode the authors call behavioral state decay, where task requirements, environment facts, and open subgoals get buried or pushed out of the context window as a long-horizon task runs. The fix is architectural, not a bigger context window or a better base model.
The memory agent runs alongside an unmodified action agent, the system that does the actual work of a terminal session or a multi-turn customer service flow. It continuously updates a structured memory bank from the recent trajectory, then makes a binary call on every step: inject a memory-grounded reminder into the action agent’s context, or stay silent. The authors describe the module as plug-and-play with frontier action agents and existing agent harnesses, meaning it attaches to a system rather than replacing one.
On Terminal-Bench 2.0, a benchmark of multi-step command-line tasks, the memory agent lifted pass@1 by 8.3 percentage points. On tau2-Bench, built around multi-turn customer-service-style agent tasks, the gain was 6.8 percentage points. Both improvements held across weaker and stronger action agents. The memory agent is not simply compensating for one model’s specific blind spots.
The paper’s ablations are the more interesting result for anyone building agent infrastructure. Selective intervention beat four alternatives:
- Passive bank exposure, where the action agent can read the memory bank on demand
- Always-on injection, which reminds constantly regardless of relevance
- Advisor-only guidance
- General retrieval
The decision of when to speak carries as much weight as what gets remembered.
That the action agent goes unmodified is the detail that matters commercially. A memory layer that attaches to an already-deployed agent is a product a team can ship next quarter. A technique that requires retraining or fine-tuning the base model is a research result, not a shipping decision. The paper’s own attempt at the harder path, training a Qwen3.5-27B model on SETA with supervised fine-tuning and GRPO, improved validation reward but achieved only partial transfer to Terminal-Bench, a gap the authors report rather than obscure.
A second agent that decides when to interrupt the first is also a second place for the system to fail. Every reminder is a judgment call about relevance, and a memory agent that speaks at the wrong moment reintroduces the exact kind of noise long-context designs were meant to eliminate. The abstract does not report the latency or token cost the memory agent adds per task, and both determine whether this pattern survives contact with a production agent that already makes multiple model calls per step.
Teams running long-horizon coding or customer-support agents that degrade over extended sessions should treat this as a wrapper worth testing before reaching for a larger context window. The mechanism targets the actual failure, state getting lost, rather than the proxy, context length.
Wu et al., “Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents,” arXiv preprint 2607.08716, posted July 9, 2026.
