A durable, shared store of how the product works — the context agents read in before they act.
The team friction Product knowledge — why this rule exists, what that flag really does, which “edge case” is actually a paying customer — lives in a few long-tenured heads. When they’re out, the team guesses; when an agent works, it has none of it and invents plausible-but-wrong behaviour. The team’s hardest-won context is its least shared asset.
The pattern The team maintains product memory as a first-class, durable artifact: domain rules, feature intent, behavioural quirks, the outcomes of past bets — curated to stay true. People consult it; agents are pointed at it as standing context. It’s the team’s long-term brain, deliberately written down instead of remembered.
Why it works with AI An agent has no history with your product — product memory is how the team gives it one. Feeding the same curated context to every agent makes their output consistent with how the product actually works, instead of with how a generic app might. Shared memory, shared behaviour.
Watch-outs Memory that drifts from reality is worse than none — assign ownership and prune. Keep it product knowledge; decision rationale belongs in ADRs, transient story context in compact temporary docs.
Related ADRs — record the why · Agent-written shared docs · Compact temporary docs