My number one frustration with LLM agents is that every session starts blank and I have to re-explain my patterns and conventions every time.
I often find myself doing things like "look at X repo and copy the patterns used for Y", which requires the repo to be on the local machine.
A couple big ones for me personally: - Auth flows - Terraform patterns for AWS
Rules and skills help to some extent, but they are difficult to synchronize across multiple agents, projects, and machines. Memory is interesting, but too noisy since it is unstructured.
So, I built AI Boost to act as a personal library with an MCP server. You tell your agent what to save (text file or public github repo at the moment. I'm working on other options including private github repos.). It packages it as a "booster", indexes it with keywords and embeddings, and next time you start a task that matches, the agent surfaces it.
At the moment: - Boosters are private by default. They are only accessible to your account. - You can publish to a community marketplace if you want, and earn credits per injection (this isn't completely ready yet, but it's coming soon). - It's an MCP server, so it should work in Cursor, Claude Code, and any client with MCP support.
Link: https://ai-boost.io (and the MCP URL is just https://mcp.ai-boost.io/mcp)
I'd love to know what you think. Especially whether the "auto-suggest before starting a task" behaviour feels useful or intrusive in practice.