Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.
The research material reports that preference violations with Claude Haiku 4.5 fell from 77% without memory to 20% with 10 lines, then rose to about 25% with longer memories. For a running spending total, a written rule still failed 44% of the time, while code that kept the total failed 0%.
The article text is unavailable in this language; an existing version is shown.
– https://t.co/j5j7F4M79V
Title: "Harness Evolution as Learning: Approximation, Generalization, and Optimization Limits of Self-Improving Personal Agents"
More memory does not keep helping personal agents, because relevant notes help up to a point and then extra lines start making the agent miss rules.
A personal agent can't learn every user preference through memory notes, and more notes eventually make it worse, so use code for anything it must count or track and keep memory short.
Written rules work for style, like signing texts with the user's first name. For a running spending total, a stated rule still failed 44% of the time, while code that kept the total failed 0%.
Memory size has a sweet spot. With Claude Haiku 4.5, violations dropped from 77% with no memory to 20% at 10 lines, then rose to about 25% with longer memories.
Agents that rewrite their own memory from user complaints improve early, then stall. The best methods ended near 48% violations, against 7.1% when the agent was simply told every preference.
View replied-to post on X