Conversation Learning

This wiki should learn from substantive project conversations. Chat is not only a delivery channel; it is also a source stream. When a conversation produces stable knowledge, the agent should write it back into the repository so later work can build on it. Sources: raw/conversations/2026-04-18-chat-learning-expectation.md, raw/patterns/2026-04-18-karpathy-llm-wiki.md.

What Counts As Durable

Write back conversation content when it contains one or more of these:

  • project decisions
  • user preferences that affect future work
  • reusable explanations or comparisons
  • definitions of important terms
  • architecture or workflow changes
  • research conclusions worth preserving

Do not create a new raw conversation source for trivial back-and-forth, ephemeral status updates, or purely local debugging chatter unless it changes future behavior. Source: raw/conversations/2026-04-18-chat-learning-expectation.md.

Default Flow

  1. Capture the durable part of the chat in raw/conversations/YYYY-MM-DD-topic.md.
  2. Update one or more wiki pages with the distilled knowledge.
  3. Update wiki/index.md.
  4. Append a conversation or synthesis entry to wiki/log.md.
  5. Run python3 tools/wiki_lint.py.

Local Rule

The default is auto-ingest, not opt-in write-back. If a future agent decides not to persist conversation knowledge, that should be because the exchange was transient, not because the user forgot to ask. Source: raw/conversations/2026-04-18-chat-learning-expectation.md.