What's this about?
The Pi team published three compact positions that circulated widely. On memory: code is the truth — you don't need a memory system or RAG for a codebase, because models are good at reading code structure directly. On tools: bash is all you need — bash composes like a programming language, so skills plus scripts usually beat MCP servers. And: build context-efficient tools. Separately, Pi's post on compaction drew a technical critique from Baoyu.
What exactly is RAG, and why would you not need it?
RAG — retrieval-augmented generation — means searching a corpus for relevant chunks and pasting them into the model's context before it answers. It became standard because models couldn't hold large document sets. Pi's argument is that a codebase is different from a pile of documents: it has structure, an entry point, imports, and a filesystem, and a model with a grep tool can navigate it the way a developer does. Building a semantic index over code discards that structure to solve a problem you can solve by just letting it look.
What's the community saying?
- "Code is the truth" is the load-bearing claim. It argues against a whole category of tooling — vector databases over source code — on the grounds that the artifact already contains its own index.
- "Bash is all you need" is a direct shot at MCP. The argument is composability: bash pipes arbitrary tools together, whereas each MCP server is a fixed interface someone has to define and maintain.
- Baoyu's critique of compaction is fair and unflattering. He read it as simply prompting the LLM to summarize while keeping the system prompt, tool calls and recent turns — lossy, with no retrieval back over history — while conceding it is "genuinely the simplest effective approach." His open question: is there any mechanism to retrieve context from earlier in the session?