Why does ClaudeCode use plain text + file system as its RAG system? Claude Code adopts a "plain text + file system" RAG design and completely abandons their original vector-based RAG solution. According to Anthropic engineer Boris Cherny, this design was not determined from the beginning, but was an active choice after verification in practice. So I did some research and discussed this topic with some group friends. I would like to briefly share my summary, purely for the sake of technical reference, without any criticism. The core concepts of the current Claude Code RAG system:
- "Do the Simple Thing First" (Do the Simple Thing First) This is the Unix philosophy that Boris Cherny has repeatedly emphasized. Claude Code is designed to be "a Unix tool, not a complex product" - the memory is the markdown file, the search is grep, and the prompt compression is simple summary. Complexity is introduced only when proven necessary.
- From "Personal Experiment" to the Origin of Product Verification According to Cherny's recollection, Claude Code originated from a personal experiment: he gave the model a bash tool and observed it writing AppleScript independently to query the music library, and thus realized that "an agent with a tool is better than a pre-retrieval context script." This insight drove the entire design direction - agent topology (multiple agents working in parallel, each with a fresh, isolated context) is better than the cumulative pollution memory of a single large agent. Earlier versions did use RAG + local vector database, but the team found that Agentic Search performed "overwhelmingly better".