Memory Center: review gates for what the agent remembers

Agent memory sounds magical right up until it remembers the wrong thing. A stale convention, a misunderstood preference, or a one-off workaround can become invisible gravity for every future session. That is not intelligence. That is drift with better branding.

MendCode treats memory as a reviewable artifact. The agent can propose what it learned, but the user decides what becomes durable. Project facts stay project-scoped. General preferences can move global only when they deserve to follow the user across repos.

Memory should have provenance

If a memory matters, it should be possible to understand why it exists. Memory Center keeps proposals, accepted entries, categories, and scopes in one place so the developer can inspect context instead of trusting a hidden vector soup.

The rule is simple: no silent permanence. A correction, a repo convention, or a repeated workflow can become memory, but not as a side effect the user never saw.

Dream is maintenance, not mythology

Dream mode exists to consolidate evidence and suggest cleanup. It is not there to invent personality. It reviews what happened, proposes useful structure, and leaves the final decision in the open.

That makes memory less spooky and more like source control: propose, review, apply, reject, revise.

Side chat keeps the main agent honest

Memory work often needs a smaller conversation: why should this be saved, what scope should it use, is the category right, does this conflict with another entry? The side chat keeps that discussion near the memory system instead of mixing it into the main coding thread.

The result is a memory model that feels practical. Less oracle, more notebook with review gates.