Code Club #3: Why AI Agents Fail at Context Management?

Another exciting edition of Code Club was held at the CroAI HQ, bringing together around 50 developers and software architects. The main theme of the evening was knowledge architecture and persistent context—specifically, what it actually takes to build AI systems that can operate reliably outside a single chat window.

The Problem with "Memory" in Modern LLMs

In a packed room, guest speaker Luka Kladarić, a seasoned cloud infrastructure expert and former CTO (Noom, Hitlist), instantly debunked a common misconception: AI agents don’t fail because the models are bad; they fail because context management is a complete mess in practice.

Although tools like ChatGPT and Claude sell "memory" as a built-in feature, most of that memory is essentially a hidden state that users cannot inspect, edit, or version-control. The models work brilliantly while you are sitting at your desk, but the system falls apart the moment you walk away because the actual context remains trapped inside an ephemeral chat session.

Key Takeaways from the Presentation and Live Demos:

• Structured Knowledge Base: Instead of relying on short-term memory within a single LLM session, Kladarić presented an approach built around a persistent knowledge system. This allows for clean context switching across multiple, completely separate workflows.

• Tools in Action: Through a live demo utilizing tools like NanoClaw, Obsidian, and Claude, Kladarić demonstrated practical examples of organizing personal and business context layers.

• Avoiding Context Rot: The discussion covered how to prevent situations where an agent loses track of the task at hand and begins mixing in irrelevant information from past interactions.

Focus on Real Engineering Problems over Hype

Kladarić, who regularly shares his strong engineering opinions on his newsletter chaosguru.substack.com, drew an audience that actively participated throughout the session. The crowd raised numerous highly technical questions regarding implementation, token optimization, and infrastructure setup.

The event once again proved that the local developer community has a strong appetite for advanced technical discussions that address real production bottlenecks, moving far away from generic marketing hype surrounding AI.

A big thank you to Luka Kladarić for an excellent talk, and to Valentina Zadrija and Herman Zvonimir Došilović for organizing and moderating yet another successful engineering meetup. See you at the next Code Club!