PMB
Local-first memory for AI coding agents
I kept hitting the same wall with AI coding agents - Claude Code, Cursor, Codex. Every new session they forgot everything: the architecture, the decisions we'd made, the "we use X, never Y" lessons. So I'd spend the first 10 minutes of every session re-explaining my own project.
I also didn't love shipping all that context to yet another cloud service. It's my code, my decisions. So I built PMB - local-first persistent memory for AI coding agents.
How it works
PMB stores your project's memory in SQLite on your own machine and feeds the right context back to the agent over MCP. No cloud, no API keys, no LLM call on the read path. One command wires it into Claude Code, Cursor, Codex, Windsurf or Zed.
A few things I'm happy with:
Hybrid recall (BM25 + vectors + entity graph), ~35 ms warm
Multilingual out of the box - a Russian query finds an English fact
A local dashboard that shows which lessons actually changed outcomes, instead of claiming "+X%"
100% offline, Apache-2.0
Where it's at
Just launched on Product Hunt, and it's listed in the official GitHub MCP Registry. Building it fully in the open.
🌐 pmbai.dev
⭐ github.com/oleksiijko/pmb
Would love your take
How do you currently deal with your agent losing context between sessions - long rules files, pasting context back in, or just living with it? And would a local memory layer be worth running, or does "one more tool" kill it for you?
2 Comments
Persistent memory feels like one of those missing layers in the AI coding workflow. Re-explaining a project every session isn't just repetitive—it changes the quality of the output because important decisions get lost. I also like that you kept it local-first. That fits the problem well.