这是一篇由原始材料转换而来的阅读页，保留了源文件的主要结构，并补充了可追溯的来源说明与链接。
Logging converts experiments from isolated events into reusable organizational memory.
Pattern: Logging
Purpose
Logging converts experiments from isolated events into reusable organizational memory.
What should be logged
At minimum, log: - run identifier - code state or commit - experiment description - metrics - resource usage if relevant - final status: keep / discard / crash
Recommended extras: - hypothesis - what changed - why it was expected to help - observed failure mode - follow-up ideas
Minimal schema example
commit metric memory_gb status description
Properties of good logs
Good logs are: - append-only or safely versioned - machine-readable - easy for humans to scan - stable across runs - concise but specific
Why freeform notes are not enough
Freeform notes help interpretation, but structured logs enable: - trend analysis - duplicate detection - automated ranking - restart and recovery - later summarization by another agent
Anti-patterns
- logging only successful runs
- changing the schema midstream without migration
- storing results only in ephemeral context
- writing descriptions too vague to be useful
Recommendation
Use structured logs for outcomes and optional side notes for interpretation. Do not make the core result dependent on conversational memory.
来源与参考
源文件： autoresearch/patterns/logging.md
来源目录： /srv/project/harness-engineering