Agent memory is useful. But memory alone does not solve the harder problem: ➤ state integrity under contradiction. Cogdex adds: • scoped beliefs • reconciliation logic • checkpointed state transitions So the agent doesn’t just “remember more.” It acts on a state you can inspect and verify. Example: the agent remembers “be concise” globally, but for repo:core/task:architecture-review the active belief becomes “go deep + technical” instead of applying the wrong style everywhere.