Skip to main content

​
Standing notes

​
In plain terms

Vector memory is a big library. These files are a sticky note on the monitor. Two notes:
  • MEMORY.md — facts about the project / environment (“deploys on Fridays”)
  • USER.md — facts about this person (“likes short answers”)
They are always injected into the prompt (when they have content). They have a hard size cap. When the box is full, the agent must shorten old lines before adding new ones. That is on purpose — it keeps the prompt small. This sits next to MemoryManager. It does not replace sessions, summaries, or vector learnings.

​
Turn it on

Agent.deep() turns this on for you. Optional caps (defaults: 2200 chars for memory, 1375 for user):
Pass storage to keep notes in SQLite/Postgres instead of RAM.

​
The memory tool

The model gets one tool named memory: target is "memory" (default) or "user". If adding would overflow:

​
What to use when

​
Deprecated overlap

The old classes still run. They will be removed in the next major version.