TypeAgent

TypeAgent is an architectural exploration for building a single personal agent with a natural-language interface, expressed as sample code. It distills LLM capabilities into logical structures for Actions, Memory, and Plans, using TypeChat schemas to define agent actions and a “Structured RAG” approach for agent memory that builds semantic ontologies over conversation history. Structured RAG demonstrates substantially better recall than classic RAG on multi-turn conversations, accurately answering questions like “what books did we discuss?” across long histories.

TypeAgent’s three principles — distill models into logical structures, use structure to control information density, and let humans, programs, and models collaborate — yield agents that handle routine tasks via pattern-based dispatch with low latency and cost, escalating to LLM inference only when needed. The result is a practical recipe for personal agents that combine the flexibility of natural language with the reliability of structured systems, and a reference for developers thinking about how to deploy long-lived agents with durable, queryable memory.