knowledge accumulation
Why knowledge accumulation matters
In Compound Engineering, knowledge accumulation is the practice of capturing what worked and what went wrong during AI-assisted development, and codifying it into the system so improvements compound over time — the reason the methodology carries its name [^45fde771-1fa4-5519-a1bd-1c6cdabd8394@438]. It is described as the hardest and most powerful part of the cycle of planning → doing work → reviewing [^45fde771-1fa4-5519-a1bd-1c6cdabd8394@438].
How it works in practice
Instead of one giant CLAUDE.md — which pollutes context with everything from evergreen code-style notes to one-off library quirks — knowledge lives in markdown files organized by category [^45fde771-1fa4-5519-a1bd-1c6cdabd8394@1212]:- brainstorms, captured via the plugin's /brainstorm command- research for post-mortems on what went wrong- solutions for best practices, developer-experience notes, and integration issuesEach file carries metadata so agent search can find it. A subagent called the Best Practices Explorer searches these files during planning and work, bringing relevant discoveries in as research without wasting your main context. Documents are never inserted wholesale into the prompt — only pulled in when relevant [^45fde771-1fa4-5519-a1bd-1c6cdabd8394@1212].
When to capture knowledge
The trigger is often simple irritation:If I get annoyed or something like “I don't want to do this again” happens, I can save it somewhere.For teams, capture findings right after the session, while the context is still fresh — not in real time, which is too slow [^45fde771-1fa4-5519-a1bd-1c6cdabd8394@1212]. Accumulation compounds personally too: the creator constantly modifies the plugin itself, adding agents and skills [^45fde771-1fa4-5519-a1bd-1c6cdabd8394@438].
- Knowledge accumulation captures lessons into the system so mistakes aren't repeated and improvements compound.- It replaces one giant CLAUDE.md with categorized markdown files plus metadata for search.- The Best Practices Explorer subagent surfaces relevant knowledge on demand, keeping main context clean.- Capture when something irritates you, or right after a session while context is fresh.