Hivemind
How it works

Memories

What Hivemind stores, and what it does with it over time.

A memory is one self-contained statement worth recalling later, a decision, a constraint, a preference, a fact about how your project works.

They correct themselves

Store something that contradicts what you said before and the older version is retired automatically. You do not end up with both, and a model does not have to guess which is current.

text
Monday:  "We cap production deploys at 3 per month"
Friday:  "We cap production deploys at 5 per month"
         → the Monday version is retired; only the newer one is retrieved

The old value is kept and readable, often it is the interesting half of the record, but it stops surfacing in results.

They earn their place

Memories you keep retrieving are promoted; ones nothing ever asks for decay. This happens on its own, driven by demonstrated use rather than by a guess about what matters.

That promotion is also what builds your profile: a memory retrieved repeatedly across threads is, by definition, a stable fact about how you work.

They are not duplicated

Send the same content twice and the second is a no-op. Give a memory your own customId and re-sending it with changed content replaces it in place, so a document you sync repeatedly does not accumulate copies.

Kinds

Kind
factA stable truth
decisionA choice and its reasoning
handoffStructured work state, see Handoffs

Kind affects how a memory is presented to a model, not whether it is found.

Forgetting

Three ways, in increasing order of bluntness:

  • Undo, straight from the notice when something is captured.
  • Delete a single memory from the dashboard.
  • Forget matching, describe what should go in plain language. This previews what it would remove by default; you have to look before confirming.