Chat Is a Bad System of Record

Conversation is ephemeral. “We decided X” in a thread doesn’t reliably reach the next teammate, the next sprint, or the next model. Agents don’t inherit your memory—they inherit whatever context you give them, which is usually too much noise or too little signal.

What works: curated, file-based context—a small set of sources of truth (briefs, ADRs, architecture notes, a context index that says what to read first). The model starts from stable ground, not from reconstructed vibes.

There’s a tension: you want the repo readable without dumping every operational artifact into it. A common pattern is splitting lightweight summaries (what agents load every time) from detailed operational history (sometimes elsewhere, still linked and intentional). That keeps product history clean without pretending specs, tasks, or client-facing material don’t exist.

The insight: if it only lives in chat, it didn’t happen—for the org, the audit, or the next run.