Core that follows you
Small rules and names you keep with every run; portable across models and tools.
Keep the Thread. AI memory and context, together.
A simple method for keeping your context and decisions in one place when working with AI. Two markdown files. Pick up where you left off, compare models on the same setup, and see receipts for every answer. Works with any model.
Receipts show who · why · source · date (plus model/version) so you can replay the same setup on any model.
🧭 Before a handoff: get the timeline.
🔀 During a pivot: see the decision & sources.
🔁 Swap models: compare on the same setup.
One trusted thread across tools and models
The problem isn't prompts, it's losing the thread. Keep a single, portable source of truth with receipts.
Receipts by default
Every change shows who/what/when/which model/source so answers are auditable.
Model-swap safe
Rebuild the same setup and compare outputs fairly as prices or policies shift.
Pick up the thread
Open a month-old topic in seconds; see what changed and where you left off.
Answer "why?" fast
Decisions and sources are one click away with timestamps and roles.
Works with any model
GPT, Claude, or whatever's next—the thread stays the same.
How the Thread works (memory + context, together)
It's one object. The Core travels with every run. The Working Set attaches when the job needs it. The Receipts timeline ties it together so you can replay the same setup on any model.
Core (always-on)
- A tiny summary of the project's nature and goals.
- Standing rules (e.g., UK spelling, cite sources)
- Names, glossary, invariants
Working Set (as-needed)
- Design/API notes, fresh research, retrieved docs
- Only what's relevant to the current task
- Droppable without losing the Core
Receipts timeline
- Each change: who · why · source · date
- Links back to the setup that produced it
- Promote important changes into the Core
Already using a proxy or context extender? The method works on top.
Related: How RAG fits the Thread
Quick answers
What is the Thread?
A single object you carry between runs: a Core you always keep, a Working Set you add when needed, and a Receipts timeline so results are provable and comparable across models.
Who is it for?
Solo builders and small teams who want their projects to remember the right things without switching tools or handing over data.
How do I get started?
Download the templates from GitHub, fill in your Core, and start logging decisions. See the Start page for a quick guide.
Do I need integrations?
No. It's just markdown files. Copy, paste, and edit them in any text editor.
How do I compare models fairly?
Replay the same setup across models so identical inputs yield comparable outputs. See same setup.
How can I prove where an AI answer came from?
Open the receipt: model & time, inputs, and linked sources. Re-run the same setup anywhere.
How does RAG fit the Thread?
RAG pulls fresh facts into your Working Set. The Core and Receipts keep what should last. Use both.