Capture then curate

Most AI memory systems fail in one of two ways. They forget everything, or they remember garbage.

Morpheus, my second brain, writes things down all day. Every routine that reads my calendar and email is allowed to capture what it finds: a new name, a project, a vendor, a promise. Capture is cheap and liberal on purpose. If it might matter, it gets written down.

But nothing goes into long-term memory directly. It all lands in a holding inbox first.

The rule this time: capture then curate.

The routines literally cannot edit the curated memory files. Read-only by design. Every so often I review the inbox together with the brain: real things get verified and folded in, noise gets deleted. A new client contact stays. A booking code from a trip that already happened gets dropped, deliberately never kept. Duplicates never make it in at all, capture is idempotent.

Why the ceremony? Because the curated files are what the brain reads on every single run. Whatever sits in memory shapes every morning brief and every answer after that. Pollute it once and the mistake repeats daily.

And stale memory is the sneaky failure. Early on, a routine reported something as still open that I had already handled. It had trusted its memory instead of re-reading the source. The fix was not more memory. It was a demotion: memory is a hint, never evidence. Anything that matters gets re-checked against the live source before the brain asserts it.

So the loop is simple. Capture everything plausible. Trust none of it until verified. Keep only what earns its place. A memory nobody prunes turns into a landfill.

Funny thing is, that is just how a good notebook works. The AI part means it now fills itself.

Got a system for deciding what your AI is allowed to remember?