Inspiration: Solving the AI Amnesia Problem

Most AI coding sessions feel like they have amnesia. You spend 20 minutes explaining your architecture, but the moment you start a new terminal window or clear the chat, that context is gone. Maxo was born from the desire to turn AI from a "chatbot" into a stateful, strategic co-creation partner. We wanted a system where the AI's "brain" lives directly inside the repository, evolving alongside the code.

How We Built It

Maxo is built as a modular extension for the Gemini CLI. At its core is a Blackboard Architecture, which decouples high-level architectural planning from tactical execution. We implemented a Python-based MCP sidecar kernel that allows Gemini to perform surgical, structured updates to a persistent memory file (maxo.md) using JSON-in-Markdown.

Key features include:

Agentic Discovery: Specialized sub-agents that use a reverse-engineered Scratchpad Pattern to map codebases top-down without hallucinations.
Asynchronous Execution Queue: Allowing a "Planner" agent to dispatch tasks to multiple "Executor" agents running in parallel across different terminal sessions.
Time-Travel Memory: Using Git integration to snapshot the AI's cognitive state at every milestone.

Challenges & Learnings

The biggest technical hurdle was managing path resolution across different user environments. We had to implement a dynamic variable system using ${extensionPath} to ensure the framework remained portable. We also learned that Structural Context Compression—representing logic as code and state as JSON—is significantly more efficient than natural language for long-term co-creation.