How to Give Claude Cowork Memory

Why does Claude Cowork forget everything?

Claude Cowork, like all large language models, works with a context window. Each session starts with a clean slate. Messages fill this window progressively, but when the session ends, everything resets. This is not a bug — it is an architectural constraint. The model has no persistent database to store your preferences or project history.

But Anthropic built mechanisms to work around this limitation. They are not always intuitive, which is why most users do not leverage them.

The CLAUDE.md file: your primary weapon

The CLAUDE.md file is the most powerful mechanism for giving Claude Cowork persistent context. It is a Markdown file that Claude reads automatically at the start of every session.

When you start a Claude Cowork session in a directory, the agent looks for a CLAUDE.md file at the project root. If found, it loads it into the initial context. Everything you write in this file will be read and considered at every session.

What to put in your CLAUDE.md

Include your tech stack, coding conventions, important decisions with dates, and things to avoid. The goal is to provide the information a new developer would need to work effectively on your project.

Three levels of CLAUDE.md

All three levels stack. Claude reads global instructions first, then project-level, then subdirectory-level.

The Projects feature

Claude Cowork includes a Projects feature that groups conversations around the same context. When you create a project, you define a working directory and reference files.

• Automatic context: project files are loaded into every new conversation.
• Per-project history: easily find conversations related to the same project.
• Reference files: attach documents Claude will consult systematically.

Global instructions: about-me.md and beyond

Beyond CLAUDE.md, you can create instruction files in ~/.claude/ loaded for every session regardless of project. An about-me.md with your experience level, language preferences, and communication style saves repetitive explanations. Rule files in ~/.claude/rules/ (typescript.md, python.md, security.md) apply thematic conventions across all projects.

Context files: skill files and beyond

For complex projects, a single CLAUDE.md is not enough. Create a structured context architecture:

Claude loads these files on demand, preserving the context window for your current task.

Advanced persistent memory techniques

Decision journal (ADR): Maintain a decisions.md file recording important architectural decisions with dates and justifications.

Session summaries: At the end of long sessions, ask Claude to generate a summary saved to sessions/YYYY-MM-DD.md. Load the latest summary at the start of your next session.

Custom slash commands: Create commands that automatically load context file sets. For example, /resume-project loading the CLAUDE.md, latest ADR, and latest session summary.

Structured memory with tags: Organize your CLAUDE.md with semantic tags that Claude can use to find relevant information quickly.

Mistakes to avoid

CLAUDE.md too long: A 500-line CLAUDE.md is counterproductive. Aim for 50-100 lines and move details to separate context files.

Contradictory instructions: Ensure your context files do not conflict. The most specific file should always prevail.

Outdated information: Update context files regularly. Add a last-updated date at the top.

Conclusion: build your memory system

Claude Cowork has no native memory, but with the right context files you can create a surprisingly effective memory system. The key is investing time in initial configuration: a good CLAUDE.md, well-structured global instructions, and clear context file organization. This investment pays off from the second session onward, when Claude immediately understands your project without you having to re-explain everything.