Hi everyone - I built Jarvis, a local-first AI memory system for markdown notes that works with Cursor.
The goal is to make notes, decisions, research, and project context in markdown actually usable inside AI workflows, without constantly pasting large chunks into chat.
Jarvis is not meant to be “more repo context” for code.
It is meant to be a memory layer for knowledge that lives outside the current codebase:
- notes
- product decisions
- research
- meeting context
- work history
- docs written outside the repo
How it works:
- markdown stays the source of truth
- notes are indexed locally with embeddings
- retrieval combines semantic search, keyword search, and graph relationships
- the memory is exposed to Cursor through MCP over stdio
- AI can not only read from that memory, but also update notes, append useful insights, and save outputs back into the workspace
What this gives Cursor users:
- less prompt stuffing
- lower token usage, because retrieval happens locally first
- better access to relevant knowledge from large markdown vaults
- a shared memory layer instead of starting from an empty chat every time
- a system that becomes more useful over time as useful outputs are written back into notes
A simple example:
instead of manually pasting meeting notes, product decisions, or old research into Cursor, Jarvis can retrieve the relevant parts from a growing markdown memory and use that as context.
I’d really appreciate honest feedback — especially whether this solves a real problem in your Cursor workflow, where the setup feels too heavy, and what would make it more useful in practice.
If the idea sounds interesting, contributions, issues, and suggestions are very welcome.