Memset.io
One brain across ChatGPT, Claude, Cursor & more.
Desktop client is so close! Until then...
Below is what I rolled out this week — mostly to close known issues, a couple of upgrades while I was in there.
Claude Code MCP authentication
If you tried connecting Claude Code to memset and saw "connected" but every remember / recall call failed with "invalid API key" — that's been fixed at the protocol level.
memset is now a full OAuth 2.1 authorization server for MCP clients. Practically, that means:
You no longer paste API keys into
.mcp.jsonfor spec-compliant clients. Just the URL — sign in through your browser, click approve, done.Claude Code's discovery flow now actually finds us. Existing configs with an API key keep working unchanged.
Cursor, Claude Desktop, VS Code MCP — all unaffected, all still work.
Stay-logged-in
Browser extension refresh token bumped from 30 → 90 days. You should hit "log back in" prompts a lot less often.
Backend now logs auth failures with enough context to debug from logs alone — no more "everything looks fine but nothing works" mysteries on my side.
Performance
Page loads outside the US should be noticeably snappier. Caught a few places where the API was telling browsers and CDNs not to cache anything safe to cache; that's now per-route.
Dashboard stats render instantly on revisit (cached locally, refreshed in background).
Direct link to the Chrome Web Store from the dashboard — one less hop to the install page.
Free tier upgrade
Ghost Memory — the autocapture from your AI conversations — is now genuinely free for everyone, no tier gate.
Polish
Fixed a sidebar bug where "Setup Guide" stayed highlighted when you clicked into its subpages.
Quieted a noisy alert that was firing on every cold-start of the API server.
If you've hit anything I haven't fixed, hit reply. Every email goes to me directly. I read all of them.
Confession: I've been weekend-warrioring on memset.io for over a year and only learned Indie Hackers was a thing yesterday. Posted about my project on LinkedIn expecting five likes from coworkers, a few people told me I should be on IH. So... here I am.
The short version: my Nana passed, I started building something in her memory, then my computer crashed and I realized... code was safe on GitHub, but everything I'd learned solving problems with AI was gone. So I built a memory layer that follows me across ChatGPT, Claude, Cursor, Gemini, all of them.
Free tier, paid tier mostly exists to stop one specific friend from melting my Cursor/Claude bill.
I genuinely have no idea what people do here besides post their products. What's worth my time as a solo weekend builder? Mostly here to learn from folks who've done this longer than me.
-
2
The context loss thing is real though, I've had Claude forget entire project setups mid-conversation and had to start over.
How does it sync across tools? Does it inject context automatically when you open a new chat or do you have to manually pull it in each time?
-
1
Great question Muhammad. Currently I have it as passive injection - i.e. the system detects what you're talking about and if a relevant memory exists within your brain it pushes that memory as an option to you; you can then decide if you want to inject that memory or if you don't want to.
Note that this only works if you're actively engaged with a system (Claude, Cursor, ChatGPT etc..). For your example, where you open a new chat, and the system has no context about what you may or may not want to investigate/explore.. you would force recollection using the /recall command.-
2
That passive injection approach is smart — less intrusive than forcing context every time. The /recall command makes sense for cold starts too.
Curious: does it work across browsers or is there an extension involved? Will give it a proper try this week.
-
1
Currently the ecosystem that feeds the brain consists of:
- Chrome Extension
- MCP
- CLI
So, it won't work across browsers.
However, I am in the final stages of building a desktop client and going through the update approvals across Windows/MAC. The desktop client would cover all surfaces on a desktop; and then I'll build out a mobile app as well.
Is a roadmap with all the future state stuff I'm piecing together on the roadmap linked off the home page.-
2
Moving to a desktop client is a massive upgrade because it takes the "brain" out of the browser silo and makes it universal across every coding environment you have open. Eliminating that browser-only limitation means the AI can actually stay synced whether you're in a terminal, a specific IDE, or just drafting documentation. By the way, I work in PR and media placement for deep-tech and productivity tools, and this vision of a "cross-surface" AI memory is a really high-authority story that would land well with major tech news outlets.
-
-
-
-
Comment