Solved my own Context Amnesia (and open sourced it)
Every time I start a new Claude/Cursor chat, I lose context. I have to re-explain my entire project.
So I built Nucleus - an MCP server that remembers for me.
The Problem:
- 5 different AI chats (Claude, Cursor, Windsurf, etc.)
- Each chat starts fresh (context amnesia)
- Re-explaining project 5+ times per day
- Losing decisions, losing architecture, losing momentum
The Solution:
- .brain/ folder (persistent context)
- Tasks (what's done, what's pending)
- Events (full audit trail)
- Sessions (save/resume context)
- Depth tracking (prevents rabbit holes and stays focused)
The Proof:
- 948 events logged (real usage, not demo)
- 4.6x productivity (312 files in 15 hours vs. 160 hours manual)
- Used daily for 6 months (dogfooding, not vaporware)
The Tech:
- Python MCP server
- Local-first (your data stays on your machine)
- Open source (MIT license)
- On PyPI (pip install mcp-server-nucleus)
The Ask:
How do you manage context across AI sessions? Do you have this problem?
PyPI Link: https://pypi.org/project/mcp-server-nucleus/
Full disclosure: I built this. Sharing to get feedback and see if others have same problem.
Love this — scratch your own itch at its finest. The .brain/ folder concept is clever. We're tackling the same "context amnesia" problem at Lumi but from the LLM memory layer side. Different angle, same enemy.
This is a great insight, and the MCP server sounds like quite an ingenious idea that solves the problem of LLM "context amnesia".
I tackled this exact problem recently, but I took a different architectural approach. Instead of a script/MCP, I built a full desktop state machine to manage the project context externally.
It’s fascinating to see us both converging on the idea that the LLM needs a manager to hold the state. If you are curious to compare how the desktop approach handles the memory persistence vs your script, I released my tool (Klyve) as a free binary on Github.
Good luck with the open source release!