Memory
The archive
How the system remembers, and what it is deliberately not allowed to remember.
An AI model forgets everything when a conversation ends. Anything it should know tomorrow has to be written down somewhere and found again.
The technique has an awkward name, RAG, which simply means: look it up first, then answer. The hard part is not storing things. It is knowing which note is current, where it came from, and what should never be stored at all.
1
shared memory for every worker, instead of one each
The layers of memory
One shared memory
A single searchable store on my own machine. Every worker reads from the same one, and every answer says which note it came from. No outside service is needed.
Notes I can read too
ObsidianA notebook that the workers write and I read on any device. If a machine keeps notes about my work, I want to be able to open them.
Handoff notes
500+After each real task: what is done, what is left, the exact next step, and what must not be redone. A stranger could finish the job from one.
The history room
The complete record is kept, but it is never loaded automatically. It is searched on purpose, and a search returns at most eight short results.
The rule book
Skills and rules live under version control, so a bad change by a worker shows up as a difference I can review and undo.
What stays out
My personal documents and my private notebook are off limits. Passwords and message contents never go into memory.
What it taught me
- 1
Two memories that can disagree are worse than one. I tested a popular memory tool beside my own, then removed it.
- 2
A remembered fact needs a source and a date. My assistant once said good night before lunch because it trusted a stale note about the time.
- 3
More memory is not better memory. Giving a worker everything makes it slower and more expensive and no smarter.