Automatically create Agent Skills from past sessions. Gemini CLI can now comb through past session data and suggest new skills based on past patterns of things you do frequently! Helps the agent self-improve. 🧠 Enable Auto Memory in /settings to try it out.

Wonder where this idea came from xD

Looks like I spend too much time working on themes/animations, designing specs for new features and resolving merge conflicts 🤣🍀

Awesome feature! I’ll give it a try. Btw, I built skill-eval for a skill related problem: having a more rigorous way to measure whether a Skill actually improves an agent’s behavior, beyond just a few examples that happened to work. For now it’s focused on Gemini CLI:

Ouuu this is super cool! Could totally see us implementing something like this to make sure the skills created are impactful enough and to discard ones that are not. 😀

i like this direction. half my repeat work is basically a missing skill definition.

hey there is something weird happening with my Gemini CLI set up lately... sometimes in the middle of a session, the model that is used somehow drops down to Gemma... and this is the screen shot I took yesterday... it started hallucinating 'Elephants' and it crashed 😁

Are you using the new experimental Gemma model routing by any chance? 🐘

no. I have Gemma installed locally through LMStudio... but that's the only connection I have with the lady.

Is there a command to promote these skills when you enter it? Mine seem cut off and unable to scroll no matter what advice I attempt from gemini. I have reverted to asking gemini to promote or discard them.

Mind explaining more what you are running into? Is it not letting you hit "Enter" on the skills in the inbox?

Sure! I hit enter on the skills on the inbox and it brings me into the skill. The skill is cut off on the bottom and no matter what I have tried, I cannot scroll down to read the entire skill. And then once in the memory there are no options visible. If I hit enter it discards.

When asked gemini I was given. This helpful description, which does not work. :)

That looks like a hallucination. I don't think those commands are the proper ones. Should just be "Enter" to select one of the options.

That looks relatively straight forward. Unfortunately I cannot get to those options. Will keep trying though. Love the system.

Is this what you are seeing? Looks like an issue with alternate buffer mode rendering... getting someone to take a look.

Hello Again. It was Terminal Buffer set to true. I can now access the options at the bottom.

When you're using the CLI, do you prefer Pro, Flash, or Auto?

I use Flash like 80% of the time personally. My workflow is to use Pro for planning and then Flash for implementation. 😎

That’s fair. I'm really excited to see what Google does with the next Flash models! The generous limits they offer across everything make them a lot of fun to use. Will the CLI always stay separate? It feels like Google has gone all-in on Antigravity while the CLI is "forgotten".

Stay tuned for Google I/O! Lots of exciting stuff being announced based on feedback from users🚀

Very excited. I just hope that if any model upgrades come, usage limits aren’t decreased. Hopefully they go up, but I doubt it 😅

Is the "agent self-improve" part also open source?

@No_Name7n11 Absolutely, the code for looking through past sessions and creating skills from them is in our GitHub repo 👍

Gemini CLI is full of cool features, and I love that. However, Gemini APIs are so unstable that I can't even test these features.

do share the telegram info lol :)

could i get it to sync with AntiGravity session? would be nice when i want to ssh into gemini cli

Love this, stoked to see it learning from past sessions. Keen to try Auto Memory, any tips for getting the best skill suggestions?

Whoa whoa whoa! This seems like an awesome feature. Gotta try this out.

Self-improving agents are the future. Mnemosyne does this for Hermes - it combs through past sessions, consolidates old memories into summaries, and helps the agent build knowledge that compounds. Been running it a month.