Wink Pings

Self-Improving Agents Must First Learn Not to Ruin Their Memory

Google Cloud's GUARD freezes the base model and lets the Agent change only prompts, skills, tool rules, and memory, using pre-execution self-checks and regression gating to decide whether an edit can stay. It reaches 80.80% on 43 hosted SDK tasks, but the gap vs. Meta-Harness is not statistically robust; what's really worth copying is the gating and memory hygiene.

Google Cloud's paper GUARD installs a regression gate on a self-improving agent harness. The base model is frozen; the agent can modify its own prompts, skills, tool rules, and memory. Whether an edit stays depends on whether its gains outweigh the behavior it breaks. arXiv DAIR.AI explainer

GUARD 论文分享图

Common problems with self-improving agents are concrete: local overfitting, catastrophic forgetting, memory contamination. The weights don't move; what changes is the harness. Reflection and scalar scores easily conflate two things: what this edit newly broke, and problems that were already unsolved. The result is fixing one spot and stepping on another, with memory getting dirtier the more it is written.

GUARD's approach is to break each task into pre-execution self-checks. Each round, these checks score trajectories, pin down rules already mastered, and rank newly broken behavior first. Candidate edits are compared using group-relative fitness, and the Soft-HURT gate lets an edit through only when gains exceed regressions. At test time, a frozen harness runs a single pass. The design isn't fancy, and the direction is clear: first let the system know what it has broken.

On the data side, 43 hosted SDK tasks reach 80.80% accuracy, 8.89 points above the initial harness, p = 0.008; 6.28 points above Meta-Harness, using 54.5% fewer tokens per task. On procedural memory tasks, update correctness is 90.1%, with zero memory contamination. An aggressive rewrite approach has close update correctness but three times the contamination. This comparison is more vivid than a single-point accuracy number: memory isn't better the more you write; if you write it dirty, you'll pay for it later.

In the replies, Karthik Bhaskar points to the caption of Table 2: all five configurations beat the baseline, 80.80% is the last of the five, and the mean over the same 43 test tasks is 77.09%; compared with Meta-Harness, p = 0.11, and the 95% CI crosses zero. So 80.80% cannot be treated as the only conclusion. Regression gating and memory hygiene are more worth copying than any single top score.

There are also several practical questions in the comments. Sam asks when harness changes get rolled back. Draxx asks whether the check set is a fixed test set or grows with new failures. Amol asks whether GUARD can catch it if an edit helps a task today but slowly changes output style, or whether it only looks at scores. These questions don't have complete answers in the thread and need the paper body. Another commenter argues that hard-coded self-improvement loops are outdated and mentions routes like Octopus-SI-OS that let the harness natively modify itself. The route can be debated, but regression testing and memory contamination still need someone to manage them.

For people building agent harnesses, GUARD's signal isn't complicated: first turn “what got broken” into an inspectable record, then talk about recursive self-improvement. Otherwise, a score bump may just crowd out old capabilities; memory looks richer, but is actually harder to use.

发布时间: 2026-11-19 23:28