A design pattern that moves AI project knowledge to a living filesystem the AI reads, writes, and maintains directly. Your projects live on your own drive, fully portable and backed up, with built-in session continuity, temporal awareness, task tracking, and multi-project coordination.
If you've used AI project features, you've seen what AI is capable of in a single conversation and felt the limitations across many. Memory that doesn't carry over reliably, no awareness of time, no ability to build and organize reference materials, no task tracking across sessions. This system was built to overcome those limitations using only native tools.
The AI gets a living directory on your local drive that it reads, writes, and maintains directly. No external tools, no code to install, no API integrations. Everything works with features already built into Claude Desktop.
The AI picks up where you left off in every new chat. It logs its own work proactively, knows what time it is and how long you've been away, and keeps startup reads lean so conversations stay responsive. Drop any files or folders in a project inbox as you work. The project checks its own inbox at startup, notices what's new, relates items to ongoing work when it can, and won't let them fall through the cracks. The AI organizes what it learns into files it maintains and indexes, reading only what's relevant to the current work instead of loading everything into context. The whole project lives on the filesystem, so it's portable to a new device or account.
The system works for a single project or many. Multiple projects can share a knowledge base and coordinate through a dedicated coordinator project.
Built with Claude Desktop (Filesystem extension, Chat, Cowork, and Code). The architecture documents, templates, and patterns are platform-agnostic and should work with any AI assistant that has filesystem access. Web and mobile work without filesystem access. The AI continues from conversational context, then writes logs and updates files when you're back on desktop.
You need the Claude Desktop app (macOS or Windows) with the Filesystem extension enabled, and a directory on your local drive for your project files.
-
Create a root directory for your projects (e.g.,
AI Projectson your Desktop). -
Enable the Filesystem extension in Claude Desktop settings and grant it access to this directory. Every project under this directory will be accessible, which is by design: it enables cross-project coordination and delegation.
-
Set tool permissions to Always Allow for the smoothest experience. The AI reads and writes files frequently during normal work, and prompting on each one adds friction. If the Filesystem extension is only scoped to your AI project directory, the AI can only touch files it manages itself, not your personal files.
-
Give Claude the architecture document and templates. Link it to this repo or download the files and provide them directly. (Some AI assistants may fetch outdated cached content from the web, so check to ensure it's getting the newest version or download it instead if you prefer.) Describe your project and ask Claude to build the workspace. It will scaffold the directory, create the files, and give you the project instructions to paste into your project settings.
That's it. Every new chat in that project reads the workspace and continues where the last one left off.
| File | What it is |
|---|---|
| workspace-architecture.md | The complete architecture. Principles, structure, file roles, startup sequence, knowledge organization. Start here. |
| templates/ | Deployable text for every file in the system. Workflow sections go into WORKFLOW.txt verbatim. Mandated file templates show the prescribed structure for handoffs, references, status files, indexes, and more. |
| patterns/ | Supporting patterns developed through use: temporal awareness, evolving state, archiving, agentic delegation. |
This is a project I maintain for my own work. Hopefully you find it useful and can adapt it to yours. If you run into problems or have suggestions, open an issue on the repo.
Built by a non-developer through iterative design and daily use.
Active development. Tested across multiple projects in different domains, continually being refined and updated.
Part of AI Project Architect — Version 4.1