Richie

Richie’s memory layer starts with deep codebase analysis across repositories and related artifacts. The goal is not just to index files, but to build a usable representation of project intent, structure, responsibilities, and technical decisions.

The pipeline is intentionally modular. Individual files are summarized first, then larger contexts are collapsed when token limits become an issue, and finally project-level summaries are produced for storage. This keeps the system cost-aware and makes large repositories tractable.

• File Analyzer Agent extracts file-level logic and responsibilities.
• Collapser Agent condenses overlong intermediate outputs without dropping key context.
• Project Analyzer Agent produces higher-level project understanding.
• Orchestrator Node coordinates sequencing and resource usage.

The resulting summaries are embedded into a vector database alongside resume text, portfolio material, and other professional context. That becomes Richie’s long-term memory.

Once the memory layer is built, Richie uses retrieval-augmented generation to answer questions in real time. The system is not a single prompt over a vector store; it uses routing logic so different kinds of questions take different paths.

• Query Classification Agent decides whether the question is document-based, general, or out-of-scope.
• Prompt Rewriting Agent clarifies vague requests before retrieval.
• Context Chatbot Agent pulls relevant memory chunks and assembles context.
• General Chatbot / Out-of-Scope Agents handle casual or irrelevant queries without wasting retrieval resources.

This structure improves relevance, avoids unnecessary token spend, and keeps the experience more stable than a single undifferentiated prompt chain.