By Teresa Torres · producttalk.org · @ttorres on X · LinkedIn
Teresa Torres explains how she stopped re-explaining her business to AI tools by giving Claude Code a persistent, file-based memory. Instead of uploading context into chats or web Projects, she stores knowledge in local files that Claude reads on demand. She argues this avoids the tedium of repeated prompting and the context degradation that comes from overloaded windows. The article introduces a three-layer memory structure: global preferences, project-specific instructions, and small reference files pulled only when relevant. The full step-by-step setup is reserved for paid subscribers.
01Key takeaways
- Store durable business context in local files so you stop re-uploading it into every chat.
- Keep global and project instruction files concise, since they load into every relevant session.
- Split reference material into small, targeted files and load only what the current task needs.
- Overloaded context degrades output quality even before hitting technical limits, so curate what Claude sees.
- Maintain memory files as your products and customers change, treating them as a compounding investment.
02Key sections
- The problem with starting from scratch
- Large language models forget everything between conversations, so users must re-supply context each time. Web Projects help somewhat but become cluttered and unwieldy as work spans multiple products.
- Why local files create memory
- Because Claude Code can read files on your machine, you can build reusable context that applies across chats and projects. Files can be mixed so Claude gets only what is relevant to each task.
- Layer 1: Global preferences
- A single CLAUDE.md file in the home directory stores how you want to work with Claude in every session, such as planning first, giving direct feedback, and asking one clarifying question at a time.
- Layer 2: Project-specific instructions
- Each project directory gets its own CLAUDE.md describing its rules, tools, conventions, and workflows, such as writing review steps or a coding stack, loaded automatically when working there.
- Layer 3: Reference context
- Detailed knowledge lives in small, separate files that Claude loads only when needed, because overloading the context window degrades output quality, a phenomenon called context rot.
03From the post
“"Can you critique the landing page for my new Story-Based Customer Interviews course?" I used to waste hours trying to get ChatGPT or Claude to adequately critique my work. I'd get frustrated by the generic feedback, the poor writing, and the suggestions that just wouldn't work for”
04Frameworks mentioned
Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.