A new “proactive memory” module lets large-language-model agents track task requirements, environment facts, and past failures—without fine-tuning the core model. In benchmark tests the addition lifted Sonnet 4.5’s score by 8.3 percentage points on Terminal-Bench 2.0 and by 6.8 points on the τ2-Bench suite; a SETA-trained memory agent raised a frozen model’s pass@1 from 37.6 % to 41.1 % on unseen problems.

Why agents lose the plot

When an LLM-driven agent follows a multi-step procedure, its internal “state” decays. Researchers call this “behavioral state decay”: the model forgets earlier instructions, key facts, or mistakes it has already made. The result is a cascade of poor decisions that abort a task long before completion.

The usual fix is to enlarge the context window—the chunk of text the model can attend to at once. That helps only until the task outgrows the window; older information is dropped and decay resumes.

A reminder that acts only when needed

The new approach adds a lightweight auxiliary memory agent that watches the main model’s reasoning trace and injects targeted reminders. Instead of pulling a static list of past snippets, the memory module decides when and what to surface, acting only when the main model appears to drift.

Because the memory learner trains separately, the primary action model stays frozen. The study shows that this separation still yields sizable gains on long-horizon benchmarks, proving that proactive reminders can compensate for a limited context window.

What the numbers say

  • Terminal-Bench 2.0: Sonnet 4.5 jumps 8.3 points above its baseline.
  • τ2-Bench suite: The same model improves by 6.8 points.
  • Pass@1 on held-out tests: A SETA-trained memory agent lifts a frozen model from 37.6 % to 41.1 %.

These lifts occur without any additional fine-tuning of the main model, so the memory component can be swapped in or out of existing deployments.

Limits of the current work

The experiments stop short of testing:

  • Scale: No evidence yet that the memory module behaves the same way with multi-billion-parameter models or tasks that run for millions of steps.
  • Production cost: Storing and retrieving reminders adds overhead, but the study does not quantify the extra memory or compute required in a real-world system.

What to watch next

Future benchmark suites will need to measure more than raw context size. Tests that explicitly track state decay and reward active reminder systems will give a clearer picture of an agent’s long-term reliability. Researchers must also show that proactive memory scales efficiently, both in model size and operational cost.

Takeaway: A small, separately trained memory module can act as a watchdog for large-language-model agents, catching and correcting drift before it derails a task.