-
Notifications
You must be signed in to change notification settings - Fork 13.6k
feat(core): background memory service for automatic skill extraction #24272
Copy link
Copy link
Labels
area/agentIssues related to Core Agent, Tools, Memory, Sub-Agents, Hooks, Agent QualityIssues related to Core Agent, Tools, Memory, Sub-Agents, Hooks, Agent Qualityworkstream-rollupLabel used to tag epics and features that are associated with one of the three primary workstreamsLabel used to tag epics and features that are associated with one of the three primary workstreams🔒 maintainer only⛔ Do not contribute. Internal roadmap item.⛔ Do not contribute. Internal roadmap item.
Metadata
Metadata
Assignees
Labels
area/agentIssues related to Core Agent, Tools, Memory, Sub-Agents, Hooks, Agent QualityIssues related to Core Agent, Tools, Memory, Sub-Agents, Hooks, Agent Qualityworkstream-rollupLabel used to tag epics and features that are associated with one of the three primary workstreamsLabel used to tag epics and features that are associated with one of the three primary workstreams🔒 maintainer only⛔ Do not contribute. Internal roadmap item.⛔ Do not contribute. Internal roadmap item.
Type
Fields
Give feedbackNo fields configured for Task.
What would you like to be added?
A background memory service that runs on session startup, scans past conversation sessions, and extracts reusable project-specific skills as SKILL.md files to
~/.gemini/memory/<project>/skills/.Key design:
read_fileto inspect promising ones~/.gemini/-only workspace, no shell access)skillsCreatedmetadataexperimentalMemoryManagerflagWhy is this needed?
Currently, knowledge learned during a session is lost when the session ends. Users repeatedly encounter the same project-specific workflows, failure modes, and conventions across sessions, but each new session starts from scratch. This feature enables the CLI to automatically learn and remember reusable workflows, reducing repetitive discovery and improving agent efficiency over time.
Additional context
This is a single-phase approach that can evolve into a richer memory system (e.g., separating per-session extraction from global consolidation).
Sub-issue of #20062.