I've been using MnemoDB on my own work for weeks now. It's the best agent memory I've used - and I built it, so calibrate accordingly. Here's what actually changed day-to-day.
It remembers. I say it once, and days later, in a session that shares nothing with the first, it still knows. Not sometimes. Every time.
Losing a session stopped hurting. This used to be the worst moment in agent work: the window dies, and so does everything the agent understood about your project. Now the context isn't in the window. It's in .memory/ in my repo, and the next agent reads it in seconds. I'm back where I was in under a minute.
I can open the files. That's the part I didn't expect to care about this much. It's plain Markdown in git. git diff shows me exactly what my agent learned or changed this week. No dashboard, no database, no vendor.
And I can tell where a memory came from. Every entry records its source - me, the agent, or a tool. Anything a tool scraped off a web page is marked untrusted and can never override something I said. That matters more than it sounds, because agents now write memory automatically, and a poisoned memory that can hijack every future session is a real attack, not a hypothetical one.
There's no setup. Install the plugin, tell your agent to remember something, and the folder appears.
It's v0.1 - early, and honest feedback matters more to me than stars. If you try it: did it ever recall something genuinely useful, did it get in your way, and did you ever actually open the files?
mnemodb.dev