Every AI session starts from zero. CRAFT’s handoff system gives Claude Cowork perfect recall across unlimited sessions — through structured files that live on your computer, not in the cloud.
Last week we launched CRAFT for Cowork — free on GitHub. This week, let’s look at the problem that started it all.
Update, July 2026: Claude’s Projects now carry per-project memory — genuinely useful, and worth using. CRAFT works a layer above it: not what Claude knows about you, but how you work, kept in plain files you own. The two work together — more on the homepage.
The Fresh Start Problem
You spend forty minutes getting Claude up to speed. You explain the project structure, the decisions you’ve already made, the naming conventions, the constraints. By the end of the session, your AI understands your work deeply. Then the session ends.
Tomorrow, you start over. Claude doesn’t remember that you chose PostgreSQL in Session 3. It doesn’t know you found a critical bug in Session 7. It has no idea that you resolved it in Session 8 and documented the fix in your lessons learned file. Every session is a blank slate, and every blank slate costs you time.
This isn’t a minor inconvenience. It’s a compounding problem. Each session that starts from zero repeats context that was already established. Decisions get revisited. Naming conventions drift. Priorities shift because the AI doesn’t know what was already tried. The longer your project runs, the worse it gets.
On a 10-session project, you might spend 30% of total time re-establishing context that already existed in a previous session. Over 50 sessions, that overhead becomes the dominant cost of working with AI. The AI isn’t slow — the amnesia is.
How CRAFT Solves It
At the end of every session, CRAFT captures a structured handoff snapshot. Not a vague summary — a complete record of everything the next session needs to pick up exactly where this one left off. Priorities, decisions, persona state, unfinished work, git status, open questions, new lessons learned.
When the next session starts, CRAFT’s initialization recipe reads that handoff file and restores full context. The AI doesn’t need to be told what happened — it reads the record and continues. Your project’s memory lives in a structured text file on your computer, not in any cloud service, not in any chat history that might disappear.
This is handled by four purpose-built recipes that work together:
Reads the latest handoff, restores persona state, checks git status, loads lessons learned, and validates the project configuration. This is the recipe that gives your AI its memory back. It includes a bootstrap mode for first-time setup so new projects initialize cleanly.
End-of-session comprehensive snapshot: accomplishments, file manifest, git status, priorities, open questions, new lessons, and persona state. Includes a validation gate and an automatic git commit so the handoff itself is version-controlled.
Specialized for moving a project between machines. A 4-phase protocol — file verification, git pull, post-pull verification, auto-initialization — ensures seamless continuation on any device. Includes fallback paths when the network isn’t available.
Lightweight in-session state preservation without the overhead of a full handoff. Designed to protect against Cowork crashes, which cause total conversation loss. Think of it as a quicksave — capture your current state before doing anything risky.
What Gets Preserved
A CRAFT handoff isn’t a paragraph of notes. It’s a structured record containing everything the next session needs: what was accomplished, what decisions were made, what files changed, what the git state looks like, what priorities are active, what questions are still open, what lessons were learned, and what the persona’s current configuration is.
This matters because AI doesn’t just need facts — it needs context. Knowing that a file was modified isn’t enough. The AI needs to know why it was modified, what constraint drove the change, and what was tried before. CRAFT handoffs capture that full picture.
CRAFT initializes sessions through two independent layers: a CLAUDE.md bootstrap file (always available, covers project fundamentals) and an active initialization recipe (dynamic, current). If either layer is stale, the other compensates — making startup robust even after framework updates.
71 Handoffs. Zero Data Loss.
CRAFT’s handoff system wasn’t developed in isolation — it was built while being used. The framework itself was produced across 71 consecutive session handoffs, each one relying on the previous handoff to restore context. Not a single handoff failed. Not a single piece of context was lost.
During that development process, the handoff system survived five context compaction events — situations where Claude’s own context window was truncated mid-session. It survived multiple Cowork crashes that destroyed the active conversation. It survived a device switch from desktop to laptop. In every case, the structured handoff file contained everything needed to continue.
The handoff system also includes a 4-layer propagation chain that ensures persistent behavioral directives — like project-specific formatting rules or communication preferences — survive not just across sessions, but across framework updates. Directives flow from CLAUDE.md through the project file, through the generator template, through the enforcement recipe. All four layers have to fail for a directive to be lost.
Why This Changes Everything
Without persistent memory, every AI session is a one-shot interaction. You can get useful work done, but you can’t build on it. Each session is isolated, and the human has to be the bridge — manually restoring context, repeating constraints, re-explaining decisions.
With CRAFT handoffs, sessions become chapters in a continuous project. The AI accumulates understanding over time. Decisions compound. Lessons learned in Session 3 inform decisions in Session 30. The human stops being a context shuttle and starts being a collaborator.
And because handoff files are structured text on your computer, you own the data. There’s no vendor lock-in, no cloud dependency, no risk of a service change erasing your project’s history. Your AI’s memory is as portable and permanent as any other file you own.
This is Week 2 of our 8-week capability spotlight. Each week we go deep on one part of CRAFT — how it works, what problems it solves, and how to use it. Follow along as we build the case for structured AI working environments.
