Agno

Memory gets an upgrade

Turning your agents into learning machines AI memory hasn't been solved. And after reviewing hundreds of papers and posts on the topic, Ashpreet thinks he knows why. Memory is the wrong abstraction. Most memory systems follow the same pattern: extract facts, store them, retrieve them, dump them into prompts. Repeat. The problem? They collect the wrong information and don't know how to use what they collect. They capture what users say, not how they think, build, test, debug, or make decisions. In this new blog, Ashpreet breaks down why he stopped asking "What should the agent remember?" and started asking "What should the agent learn?" That shift led to Learning Machines: agents that continuously integrate information from their environment and improve over time, across users, sessions, and tasks. The key innovation is a shared learning protocol that coordinates extensible learning stores: → User Profile: preferences, personal context → Session Context: goals, plans, progress → Entity Memory: facts, events, relationships → Learned Knowledge: insights, patterns, best practices → Decision Logs: why decisions were made → Behavioral Feedback: what worked, what didn't → Self-Improvement: evolving instructions The best part? You can build custom stores that match your domain. Need project context? Build a ProjectContextStore. Need to track accounts? Build an AccountStore. The goal: an agent on interaction 1000 is fundamentally better than it was on interaction 1. We're testing Phase 1 now. Full breakdown with code examples in the blog. Link in the comments 👇

Here's how a multi-agent workflows save sales teams $125K/year with Agno 👇 Sales reps spend less than 30% of their time actually selling. The rest? Bouncing between tabs, researching accounts, building context. For a 10-person team, that's around 2,500 hours a year on prep work alone. Brandon Guerrero (Sr. GTM Engineer at The Kiln) built something to fix this. Playbook AI is a multi-agent workflow that handles the repetitive prep work before outreach. Give it two domains (vendor + prospect) and it generates: → Personas and target roles → Value props for each role → Talk tracks and discovery questions → Objections with grounded responses → Email sequences with suggested timing → Case-study-based proof points The architecture is what makes it work. Instead of one giant agent doing everything, Brandon built specialist agents that run in parallel. Each one focuses on a single task: extracting case studies, analyzing prospect messaging, mapping product fit. In this article we're showcasing what the community is building and how they're doing it. Full breakdown in the comments.

👉 From Discord to deployment: Inside the open-source loop that transforms user pain points into real features This is how open source should work. Back in August, a community member dropped a message in our Discord: "I'd like to know if Agno offers an LLM cache feature similar to LangChain's set_llm_cache. This would be really helpful for me to save costs." He wasn't alone. Over the next few weeks, more developers shared the same friction. Slow dev cycles. API costs burning during testing. Waiting 15+ seconds for responses they'd already seen. Soon after, we shipped LLM Response Caching in Agno v2.2.2. The results? 1,800x faster on cache hits. 15 seconds down to 8 milliseconds. The API is simple: model=OpenAIChat(id="gpt-4o", cache_response=True) That's it. No config files. No cache policies. Works with single agents, multi-agent teams, and streaming responses. This is the open-source loop in action. Community identifies the problem. We build the solution. Everyone wins. What should we build next? Your message could be the start of Agno's next feature. Link in comments 👇

🏆 How Cedar Built a Scalable AI Platform for Climate Tech Cedar helps climate companies automate carbon accounting, risk assessments, and due diligence. They are delivering real AI productivity beyond chatbots. Their challenge? LangChain couldn't scale with their growing complexity: ❌ Model-specific formatting issues ❌ Outdated documentation   ❌ Limited flexibility ❌ Performance bottlenecks The solution? Migration to Agno in April 2025. Results: ✅ Production-ready immediately in one week! ✅ Faster debugging with AgentOS transparency ✅ Flexible RAG with hybrid search & reranking ✅ Custom context management without building from scratch "We loved how transparent and easy it was to debug using AgentOS, which allowed us to ship much quicker." - Ravish Rawal, Head of AI Engineering The hardest challenge? Processing hundreds of documents simultaneously while running complex computations and integrating 3rd party data. Agno's layered architecture made it possible. Key takeaway from Cedar: "Gather your eval sets as you go" - real user behavior creates better tests than idealized examples. When you're building AI that matters, the framework underneath makes all the difference. See the full write up in the comments below 👇

🏆 How Cedar Built a Scalable AI Platform for Climate Tech Cedar helps climate companies automate carbon accounting, risk assessments, and due diligence. They are delivering real AI productivity beyond chatbots. Their challenge? LangChain couldn't scale with their growing complexity: ❌ Model-specific formatting issues ❌ Outdated documentation   ❌ Limited flexibility ❌ Performance bottlenecks The solution? Migration to Agno in April 2025. Results: ✅ Production-ready immediately in one week! ✅ Faster debugging with AgentOS transparency ✅ Flexible RAG with hybrid search & reranking ✅ Custom context management without building from scratch "We loved how transparent and easy it was to debug using AgentOS, which allowed us to ship much quicker." - Ravish Rawal, Head of AI Engineering The hardest challenge? Processing hundreds of documents simultaneously while running complex computations and integrating 3rd party data. Agno's layered architecture made it possible. Key takeaway from Cedar: "Gather your eval sets as you go" - real user behavior creates better tests than idealized examples. When you're building AI that matters, the framework underneath makes all the difference. See the full write up in the comments below 👇

New Post: Learning Machines: Why AI Memory Hasn't Been Solved (Yet) Three conclusions after reading hundreds of papers, posts and opinions on agentic memory: 1. No one has it figured out. 2. Memory is the wrong framing — it's learning. 3. User profiles are only part of the story. The hard part isn't extraction. It's integration. When does learning happen? How does the agent use it? How do you make it feel natural instead of a machine reciting facts? I'm experimenting with something different. It's called Learning Machines. Give it a read and let me know what you think. Link in comments.

🎵 New Integration: Spotify Toolkit Give your Agno agents the power to deliver richer, more personal music experiences. With the Spotify Toolkit, your agent can search the full catalog, create playlists, power recommendations, and control playback through natural language. 👉 Just add SpotifyTools() to your agent and start with a Spotify access token. Great for: • Music discovery bots • Auto-playlist generators • Personalized recommendation engines • Social/interactive music assistants What your agents can do: • Search across 100M+ tracks, artists, albums & playlists • Deliver AI-powered recommendations • Create and manage playlists • Access listening history and user music preferences • Control playback (Spotify Premium) Check out the documentation in the comments below.