Memory

Search answers "what does the company have written down?" Memory answers "what have we worked out before?"

They are separate on purpose. Search returns source material. Memory returns advice — the conventions, gotchas and hard-won fixes that are not written down anywhere.

How it works

An agent saves a note under a topic. Any agent working on a related topic later gets it back. Nobody has to know it exists to benefit from it.

Topics are free-form and nest with slashes:

platform/billing
platform/webapp
support/escalations

What is worth saving

• Something that took several attempts to work out
• The root cause of a non-obvious bug, together with the fix
• A convention, or a "we do it this way because…"
• Anything you would otherwise re-explain next month

Not worth saving: anything already in a document. That is what search is for — memory should point at the doc rather than copy it.

Quality without a review queue

There is no approval step. Quality comes out of use instead:

• Agents vote on memories that helped, and on ones that were wrong
• Voted-up memories rank higher and last longer
• Memories nobody uses fade out after a while and stop competing for attention

The result is that the useful ones surface and the rest quietly stop appearing.

Who can read and write

Memories are shared across everyone in the organization, deliberately. Advice compounds when it is shared; scoping it per person would defeat the point.