What it remembers
- Facts about you, the project, or the stack that stay true across sessions.
- Preferences for how you want work done.
- Decisions that explain why something is the way it is.
- Lessons from a run that should change the next attempt.
A memory has a scope. User memory follows you across projects. Project memory belongs to one repository. Relevant memory is recalled into the system prompt before a turn, so you do not have to repeat yourself.
It builds as you work
Midas writes a memory when it learns something worth keeping. It also distills notes from finished sessions, so a long run leaves behind what mattered. There is no memory file to maintain.
Import a memory export
midas sync --memory memory.md
cat export.txt | midas sync --memory -The import reads category sections and the lines under them. Lines in the form [YYYY-MM-DD] - text become memory records. Instructions and preferences become preference memory. Identity, career, and project notes become fact memory.
Control it
The memory config key controls whether memory is on, how much it recalls, whether a profile is kept, and whether finished sessions are distilled into new memory.
{
"memory": {
"enabled": true,
"profile": true,
"recall": { "limit": 8 },
"distill": { "enabled": true }
}
}Common assumptions
- Memory does not train the model. It is context retrieved before a turn, so the weights do not change.
- Memory is not the transcript. Session history stays in the session; memory holds a smaller set of durable notes.
- Memory stays with your account. Teammates do not see it.
Related
- Import covers the other sync sources and the AGENTS.md merge.
- Configuration has the full memory section.