Hey everyone,
I was frustrated by Cursor losing track of architectural decisions between sessions and re-introducing old bugs.
Existing memory tools either call slow cloud vector DBs (which breaks NDA on client code) or rely on the LLM remembering to invoke a tool mid-thought.
I built **GRM** — a 100% local, model-free graph memory written in Rust:
• **Zero-dependency 0.5MB binary** (No Docker, no Python, no OpenAI embeddings).
• **Intrinsic pre-flight injection**: Context arrives *before* the agent reasons.
• **12ms RAM-resident daemon** (PMI word co-occurrence graph + 2-hop spreading activation).
Tested over 66 continuous days of coding (24,625 memories held, 56.7% measured hit rate).
**Live demo & 7-day free trial (no signup/card needed):** https://grm-jade.vercel.app/
To test in 30 seconds:
1. Download the 0.5MB binary
2. Run `.\pm.exe install`
3. Restart Cursor and ask it about past project code.
Would love feedback on how it behaves on your repositories!