“You’re not using conversations as a database. You’re using it as a focused conversation around a particular set of context. Mentally separate the conversation from the library of documents related to your project.” — Lou
Session context: 2026-05-28_Mastermind — Joanna was stuffing 150-page legal documents into her chat threads and hitting memory limits; Lou diagnosed the real problem as a category error about what a conversation is for.
Core Idea
Most people treat an AI chat like a filing cabinet: they dump every document into the conversation so the model “has everything.” Lou’s correction is that this is exactly backwards. The conversation has a small, fixed memory; the moment you fill it with raw files you’ve “filled it up with content and didn’t leave enough room for intelligence.” Worse, most of those files are irrelevant to any single question — so they don’t just crowd the context, they actively degrade the answers by surrounding the relevant signal with noise.
The fix is a clean separation of two things people conflate. The library is every document related to your project, sitting on disk in a folder — ten files or a thousand, it doesn’t matter, because they’re not in memory. The conversation is a focused exchange about one issue, into which Claude pulls only the files it needs to answer the question in front of it. A CLAUDE.md file at the top of the folder is the index: it tells Claude what each file contains (“the purchase agreement is here, the legal claim is in legal-claim.md, the case context is in this paragraph”) so the model knows what exists without loading it. When you ask a question, Claude checks the index, decides which two or three files it needs, reads those, and answers — managing its own context in and out instead of you doing it manually.
This is why a folder-based workflow (Claude Code) beats dumping files into a chat or even a cloud project: the library is durable. When a conversation gets long and noisy, you don’t lose anything — you start a fresh thread, point it at the same folder, tell it to read CLAUDE.md, and you’re instantly back to full context with none of the accumulated back-and-forth. The folder is the memory; the conversation is disposable.
Practical Application
Set up any document-heavy project this way before you start chatting:
- Make one folder for the project. Drop in every relevant document — don’t organize it cleverly, just put it there.
- Create a
CLAUDE.mdat the top that acts as an index: one line per file describing what it contains, plus a short “case context” paragraph (or a pointer to acontext.mdif it’s long). - Run Claude from that folder. Ask one focused question at a time. Let Claude decide which files to open.
- When the thread gets long or starts circling, start a new one — point it at the folder, say “read
CLAUDE.md,” and keep going. Never re-upload.
The test of whether you’ve internalized this: if you ever find yourself pasting a file into the chat “so it remembers,” stop — put it in the folder and tell CLAUDE.md where it is.
Related Insights
- Insight - Design AI Systems for Maximum Composability and Minimum Context Pollution — the architect-level version of the same principle; this is the user-level entry point to it.
- Insight - Control AI Reasoning Effort to Stop Context Pollution — context pollution as the failure mode this separation prevents.
- Insight - The Resolver Pattern — Your CLAUDE.md Is a Pointer File, Not a Knowledge Store — files-on-disk as the durable substrate that survives any platform.
- Insight - Ship the Folder, Not the Polished Article — Your R&D Trail Is the Deliverable — once the folder is the memory, the folder itself becomes the deliverable.
Evolution Across Sessions
Builds on Insight - Design AI Systems for Maximum Composability and Minimum Context Pollution (2026-03-12), which established composability and context-pollution avoidance as an architecture concern. New development: this session reframes the same principle as a beginner’s mental model — “separate the conversation from the library” — and ties it concretely to CLAUDE.md-as-index and starting fresh threads against a persistent folder. It’s the on-ramp version of an idea the vault previously only stated at the systems-design altitude.