Let every failure grow into the next verifiable rule

Experience enters system capability only after extraction, verification and a reversible release.

Select the experience, verify the change independently, then release it gradually.

01 / SIGNALSExtract repeatable failures first

Each run leaves successes, failures, human corrections and environmental changes. These can become learning signals, but unfiltered traces also fossilize accidental mistakes into long-term rules. The goal is not more logs; it is recognizing recurring decision gaps, tool errors and corrections. One unusual failure must not become a global policy.

02 / UPDATE LAYERSChoose the smallest update layer

Experience can enter knowledge, instructions, programs or parameters, with cost and risk rising at every layer. Knowledge adds facts, cases, indexes and freshness. Instructions update procedure, boundaries, examples and stopping conditions. Programs turn stable experience into tests, tools or validators. Parameters belong only after data and evaluation mature. What a knowledge base can solve does not need a prompt change; what a rule can constrain does not need model training.

03 / VERIFY & RELEASEVerify a candidate independently before release

New knowledge may be stale, rules may conflict and programs can damage old work. A continuous-learning system needs shadow tests, version records, staged rollout and fast rollback. A proposed change should face tasks separate from the trajectory that inspired it, otherwise the system only proves that it memorized the incident.

04 / FORGET SAFELYForget safely to keep capability current

Long-term learning also includes deletion. Remove obsolete facts, low-quality memories and strategies that no longer apply, while preventing sensitive information from entering durable assets. “Sleep learning” is closer to periodic sorting, consolidation, verification and capability maintenance than automatic self-rewriting. First exercise: collect 30 failed trajectories, choose only one repeated problem for a candidate rule, then run old tasks as regression tests for side effects.

Sources

bojieli/ai-agent-book.
Chapter 8 — Continuous learning.
This is an applied reorganization, not a replacement for the source book; use the source text and code for exact claims.

Next: multimodality and real-time interaction.

Continue to multimodality

Let every failure grow into the next verifiable rule.