Lydia Hallie posted on X that subagents can be given persistent memory through a "memory" field, arguing that a subagent does not inherit the main session’s auto memory and instead "forgets everything between runs." In her description, the field gives each subagent its "own dir across sessions," with memory loading before execution and writing back for the next run.
Hallie also outlined three memory scopes: "user", for learnings that can carry across projects; "project", for repo-specific memory that she recommends committing to git so teammates’ agents can use it; and "local", for notes that stay in the same repo but remain out of git.
In a follow-up reply, Hallie addressed a question about combining user and project memory, stating that this "can't" be done with the memory field alone. She suggested instead setting memory to "project" and adding instructions such as "Also read <path to user prefs>...". She also pointed to the documentation in a separate post: Docs here!
Source: Lydia Hallie on X
