Gemini CLI: Memory Inbox for Skill Review and Patching
Gemini CLIGemini CLI v0.39.0 introduces the /memory inbox command, giving users a dedicated interface to review, approve, edit, or discard AI-extracted skills before they are promoted to the persistent skill library. The Background Memory Service, launched in v0.38.0, now has a complete feedback loop: the service automatically mines past sessions and drafts candidate skills, while /memory inbox puts users in full control of what gets retained. Google also added skill patching support in this release, enabling the system to propose targeted edits to existing skills rather than only creating new ones.
Sources & Mentions
5 external resources covering this update
The Memory Inbox: Closing the Loop on Automatic Skill Extraction
Gemini CLI v0.39.0, released on April 23, 2026, delivers the final piece of the Background Memory Service introduced in v0.38.0: a dedicated inbox for reviewing, approving, and patching AI-extracted skills.
How the Background Memory Service Works
The Background Memory Service runs silently at session startup, scanning past Gemini CLI conversations for recurring workflows. A specialized extraction sub-agent — designed to be conservative by default — identifies multi-step procedures that appear repeatedly and drafts them as candidate SKILL.md files. Sessions must be idle for at least three hours and contain 10 or more user messages before they are considered for extraction. A lock file prevents simultaneous extraction runs across multiple CLI instances.
When the service completes a run that produced new candidates, Gemini CLI surfaces an inline notification indicating how many items are waiting in the inbox.
The /memory inbox Command
The new /memory inbox command opens a review dialog listing each draft skill with its name, description, and source sessions. From this interface, users can:
- Read the full skill content before making a decision
- Promote skills to the user-level or workspace-level skills directory
- Discard unwanted drafts
- Accept or reject patch proposals for existing skills
Promoted skills become discoverable in the next session and follow the standard skill precedence hierarchy. Skills not promoted remain on disk in the inbox directory and can be reviewed at any time.
Skill Patching Support
Beyond creating entirely new skills, v0.39.0 adds skill patching support. The extraction sub-agent can now propose targeted modifications to skills that already exist in the library — for example, updating a workflow step that has changed or adding a new variant. Proposed patches are dry-run tested before they reach the inbox; invalid patches are automatically discarded, so users only see viable edits.
Plan Mode: Confirmation for Skill Activation
In Plan Mode, any attempt by an agent to call activate_skill now requires explicit user confirmation before proceeding. This closes a potential trust boundary gap where Plan Mode could silently load new skills without the user's knowledge, bringing skill activation in line with the existing confirmation patterns for file writes and shell commands.
UI and Input Improvements
Version 0.39.0 also ships several quality-of-life improvements. Mouse clicking is now supported for cursor positioning in multi-line AskUser prompts — users can click anywhere in the text area to reposition the cursor rather than relying solely on arrow keys. Scroll momentum has been enhanced for smoother navigation through long agent outputs. File paths are correctly displayed again in edit and write tool confirmation dialogs, restoring a regression from a previous release.
Reliability and Architecture
Under the hood, v0.39.0 introduces a decoupled ContextManager and Sidecar architecture that separates context management from the main agent loop, improving stability for long-running sessions and multi-agent workflows. Several memory leaks in subagent lifecycle management and MCP subprocess handling have been resolved, along with a fix for PTY exhaustion in high-volume shell workloads.