Every AI developer knows this feeling: you spend hours refining a prompt, teaching the model context, only to start the next session with a blank slate. Context windows keep getting bigger, but they never feel big enough.
RAG helps with retrieval, but it doesn't answer "where did I leave off last time?"
Existing Approaches
Most solutions attack the problem from the same angle: hold more stuff.
- Bigger context windows — more cost, more noise.
- Full state persistence — expensive and drags in stuff that doesn't matter.
- RAG — great for knowledge bases, useless for "what was I working on three sessions ago?"
None of them solve what information actually matters.
The Insight
What if we stopped asking "how much can we remember?" and started asking "what's worth remembering?"
Most of what happens in a session is throwaway. The stuff that actually matters is a much shorter list:
- Lessons learned
- Key decisions
- Code facts
- Where you left off
That's a way smaller problem than "remember everything."
Agent Memory Protocol (AMP)
I designed AMP as a pattern, not a product. Three primitives:
- Checkpoints — Where was I? What's done, what's next?
- Lessons — What did I learn that I don't want to learn again?
- Memory — Facts about code, systems, and context worth keeping.
The protocol defines the interface and gets out of the way.
Nellie: My Implementation
My implementation is Nellie — a Rust-based MCP server that runs locally. No cloud, no subscription.
Before Nellie I burned hours every week to context reset. Now I pick up where I left off. Lessons compound. I stop repeating the same mistakes.
The Point
Context reset is a symptom. The fix isn't bigger buckets — it's being smarter about what goes in them.
AMP is open. Implement it however makes sense for your stack.