Subagents do not inherit the parent session's auto memory and normally forget everything between runs. The new memory field gives each subagent its own persistent directory: it loads state before starting and writes back after finishing, so learnings carry forward across sessions.
The Team plan minimum dropped from 5 seats to 2, making shared projects, admin controls, centralized billing, SSO, and enterprise search accessible to smaller teams.
Some Max 5x plan users saw Fable 5 running on credits-only instead of their subscription entitlement. The fix was already propagating; restarting Claude Code resolves it.
The memory field for subagents supports three scopes: user (cross-project), project (repo-specific, committable to git so teammates share it), and local (repo-specific but gitignored, for private notes).
A link to the official documentation for the three subagent memory scopes (user, project, local) described in the preceding thread.
The memory field alone only supports one scope at a time, but you can layer scopes by setting memory to 'project' and adding explicit instructions pointing the agent to a separate user-prefs file path.
In response to a note that MEMORY.md is easy to inspect and edit, a developer raises that persistent memory for subagents also needs provenance, expiry, schema versioning, and a reset mechanism. Without those guardrails, a single bad assumption can quietly propagate across every future run.
A bug that briefly affected Claude Code was live for only a few minutes before a fix began propagating. The team member caught it on their personal account during late-night coding.
In response to a user still experiencing an unresolved problem after a restart, the suggestion is to check that the latest version is installed.
