Gemini CLI v0.40.0: Tiered Memory, Gemma and streamlined UX

[jackwotherspoon]

Here is Gemini CLI’s April 28th weekly update for v0.40.0

This new release has over 150 improvements, focusing on hardened security, a new tiered memory system, local model support for Gemma (routing only for now), and significant performance optimizations.

🚀 New Capabilities

• Local Gemma for Model Routing (Experimental): Initial support for local Gemma models, currently used to power intelligent model routing. The team is working toward full local model execution in future updates. Use gemini gemma to get set up and /gemma in session to check your status. (feat(cli): add streamlined gemini gemma local model setup #25498, @Samee24)
  • To enable: Use the /settings dialog or add "experimental": { "gemma": true } to your settings.json.
• Task Tracker (Experimental): Maintain an internal, persistent graph of tasks so Gemini CLI can better monitor progress on complex objectives. (docs(tracker): introduce experimental task tracker feature #24556, @anj-s)
  • To enable: Add "experimental": { "taskTracker": true } to your settings.json.
• Compact Tool Output: Tool outputs (file reads, search text, list directory) are now displayed in a cleaner, structured format for better readability. (@jwhelangoog)
• Topic Narrations: Gemini CLI now outputs topics which act as headings so you can quickly see what the agent is up to instead of the agent spamming “I’ll do this…”. Topic updates are now narrated by default and more descriptive across all agents. (feat(core): enable topic update narration by default and promote to general #25586, @gundermanc)
• MCP Resource Support: Finalized support for Model Context Protocol (MCP) resources. (feat(core): add tools to list and read MCP resources #25395, @ruomeng)

🧠 Smarter Memory & Context

• Tiered Memory System: A complete rebuild of the agent's memory architecture. The agent now persists project context by directly editing Markdown files across four distinct layers, replacing the legacy save_memory tool with a more transparent and flexible approach:
  • Project: Team-shared conventions, architecture rules, and workflows stored in ./GEMINI.md (committed to your repo).
  • Subdirectory: Highly specific overrides or instructions scoped to a part of your project (e.g., ./src/GEMINI.md).
  • Private: Project-specific personal notes and machine-local setups stored in a private MEMORY.md index (not committed to the repo).
  • Global: Cross-project personal preferences and coding styles that follow you into every workspace, stored in ~/.gemini/GEMINI.md. (refactor(memory): replace MemoryManagerAgent with prompt-driven memory editing across four tiers #25716, @SandyTao520)
• Auto Memory (Experimental): Automatically extract reusable skills from your past sessions. This background service combs through idle session transcripts to identify recurring procedural patterns and hard-won fixes.
  • How it works: A specialized extraction agent analyzes your local chat history (focusing on sessions with 10+ messages) to find common workflows. It then drafts these patterns as new SKILL.md files or patches to existing ones.
  • Review Inbox: Use the new /memory inbox command to review these drafts. You can inspect the logic, see which sessions inspired the skill, and choose to promote it to your permanent workspace or global skills library.
  • To enable: Use the /settings dialog or add "experimental": { "autoMemory": true } to your settings.json. (feat(memory): add /memory inbox command for reviewing extracted skills #24544, feat(config): split memoryManager flag into autoMemory #25601, feat(core): add skill patching support with /memory inbox integration #25148, @SandyTao520)
• Real-time Suggestions: The @ file recommendation system now uses real-time watchers to detect new files. (feat: detect new files in @ recommendations with watcher based updates #25256, @prassamin)
• Session Persistence: Added support for session persistence and JSONL-based streaming for chat recordings. (Fix/allow for session persistence #25176, @ahsanfarooq210; feat(core): migrate chat recording to JSONL streaming #23749, @spencer426)

🛡️ Security & Trust

• Hardened Command Execution: Stricter validation for shell commands and a new core tools allowlist to keep your environment safe. (feat(core): enhance shell command validation and add core tools allowlist #25720, @galz10; Fix/command injection shell #24170, @famousrajbhatt)
• Secure Workspace Loading: More secure handling of project configurations (.env files) and enforced trust in headless/IDE environments. (feat(cli): secure .env loading and enforce workspace trust in headless mode #25814, @ehedlund; Fix: Disallow overriding IDE stdio via workspace .env (RCE) #25022, @M0nd0R)
• Sandbox Protections: Improved tool level controls, custom profiles, and centralized path utilities for restricted environments. (feat(sandbox): resolve custom seatbelt profiles from $HOME/.gemini first #25427, @mvanhorn; fix(core): allow explicit write permissions to override governance file protections in sandboxes #25338, @galz10; refactor(core): extract and centralize sandbox path utilities #25305, @ehedlund)
• Credential Safety: Enhanced keychain fallback logging and secure storage refinements. (fix(core): use debug level for keychain fallback logging #25398, @ehedlund; chore: switch from keytar to @github/keytar #25143, @cocosheng)
  refinements. (fix(core): use debug level for keychain fallback logging #25398, @ehedlund; chore: switch from keytar to @github/keytar #25143, @cocosheng-g)

🛠️ UX Improvements

• New Commands: Added /new as a quick alias for starting fresh sessions. (feat: add /new as alias for /clear and refine command description #17865, @vedantmahajan)
• Better Accessibility: New colorblind-friendly themes and improved theme rendering. (feat: add github colorblind themes #15504, @Z1xus; fix(cli): ensure theme dialog labels are rendered for all themes #24599, @JayadityaGit)
• Native Notifications: Integrated OSC 777 for native terminal notifications. (Use OSC 777 for terminal notifications #25300, @jackyliuxx)
• Interaction Refinement: Mouse-clicking support for multi-line answers and improved scroll momentum. (feat(cli): enable mouse clicking for cursor positioning in AskUser multi-line answers #24630, @Adib234; feat(ui): added enhancements to scroll momentum #24447, @devrandalpura)

🐞 Performance & Reliability

• Memory & Speed: Major optimizations to startup speed and large output stream handling. (fix(devtools): reduce memory usage and defer connection #24496, @SandyTao520; fix(cli): optimize startup with lightweight parent process #24667, @sehoon38; fix(core): remove buffer slice to prevent OOM on large output streams #25094, @spencer426)
• Offline Readiness: Bundled ripgrep binaries for better offline performance. (feat: bundle ripgrep binaries into SEA for offline support #25342, @scidomino)
• Fixes: Resolved issues with heredocs, line endings in ignore files, terminal color detection, and IDE integration. (fix(cli): use newline in shell command wrapping to avoid breaking heredocs #25537, @cocosheng-g; fix(core): handle line endings in ignore file parsing #23895, @cynthialong0-0; fix(core): detect kmscon terminal as supporting true color #25282, @Adib234; fix(core): use dynamic CLI version for IDE client instead of hardcoded '1.0.0' #24414, @thekishandev)

🤝 Contributors

A huge thank you to our 40+ contributors who made this release possible!
@SandyTao520, @ehedlund, @galz10, @gundermanc, @jasonmatthewsuhari, @mahimashanware, @JayadityaGit, @g-samroberts, @scidomino, @ruomengz, @vedantmahajan, @jackyliuxx, @Z1xus, @Samee24, @M0nd0R, @prassamin, @cocosheng-g, @xoma-zver, @thekishandev, @mini2s, @famousrajbhatt, @devrandalpura, @adamfweidman, @euxaristia, @DKbyo, @ahsanfarooq210, @gordonhwc, @cynthialong0-0, @anj-s, @sehoon38, @mvanhorn, @NicolasO, @abhipatel12, @chrisjcthomas, @rcleveng, @spencer426, @clay, @Adib234, @joshualitt, @kevin8093, @jerop, @TanmayVartak, @jacob314, @mbleigh, @anjaligarhwal1610, @jesse.rosenstock, @markgriffith, @AbhijitBalaji, @mjunaidshaukat, @junaiddshaukat, @aishaneeshah, @kimjune01, @sripasg, @yunaseoul, @chernistry, @jwhelangoog, @dogukanozen, @AdamyaSingh7

[yortem]

It would be nice if I could use any of these new improvements but it says "Thinking..." for the last 2 hours.

[AcnoLite]

Same here