Gowry

Building afair (memory across every AI you use) and other things at the edges: energy, health, AI tooling. Made in Germany.

[Gowry @gowry]

https://t.co/DcI2xiEJ3A

[Gowry @gowry]

Two weeks since I open-sourced afair, one memory across every AI you use.

What shipped since:

>> You can now bring your existing history in. Point it at your ChatGPT, Claude, Obsidian, or Notion export and it becomes memory your other tools can read. No more starting from an empty vault.

>> The memory also keeps itself current now. It doesn’t just store and read back. It finds the recurring threads in what you’ve told it and maintains a live, cited synthesis of each, and updates them as things change. You never tag or file anything.

>> Plus a read-only Memory Mirror so you can see what your vault actually holds.

[Gowry @gowry]

@pmarca Best Family car I’ve ever owned 😭

[Gowry @gowry]

@elonmusk Compared to him, I was way more active, and he had 7x reach compared to his followers, and I got not even 1/10th when I announced my open-source project afair.

Not sure it this applies to everyone here on x.

[Gowry @gowry]

dear @claudeai and @AnthropicAI team, I see every day a few new comments on this issue acknowledging the non fixed issue

https://t.co/114gsWzolU

[Gowry @gowry]

It seems like, that custom remote MCP connectors are broken on @claudeai and the @claudeai Desktop app right now. Mine (afair, my open-source MCP memory server) shows "Connected" with all 3 tools in Settings, but the assistant in chat never gets them. Same server works fine in Claude Code and other harnesses.

It's not my server: my logs show your own broker (UA "Claude-User") completing the full handshake against my endpoint, with tools/list and tool calls all returning 200s, yet the tools still never reach the model. So it's the web surfacing layer, not auth.

Not just me either. Same stack and symptom, open since Jun 21 with no reply: https://t.co/v02ppn6L8Y (plus the #470 to 476 cluster). Started mid-June, right around the connector changes. Is a fix or rollback coming

@AnthropicAI @bcherny @ClaudeDevs @claude_code

[Gowry @gowry]

It seems like, that custom remote MCP connectors are broken on @claudeai and the @claudeai Desktop app right now. Mine (afair, my open-source MCP memory server) shows "Connected" with all 3 tools in Settings, but the assistant in chat never gets them. Same server works fine in Claude Code and other harnesses.

It's not my server: my logs show your own broker (UA "Claude-User") completing the full handshake against my endpoint, with tools/list and tool calls all returning 200s, yet the tools still never reach the model. So it's the web surfacing layer, not auth.

Not just me either. Same stack and symptom, open since Jun 21 with no reply: https://t.co/v02ppn6L8Y (plus the #470 to 476 cluster). Started mid-June, right around the connector changes. Is a fix or rollback coming

@AnthropicAI @bcherny @ClaudeDevs @claude_code

[Gowry @gowry]

Since going open source, afair's memory got a trust layer: it knows how sure it is of each fact, it stops overwriting your corrections, and it never silently loses a memory.

[Gowry @gowry]

Good question, and it's exactly the axis. Default recall returns the current interpretation; the superseded one isn't deleted (an append-only substrate), so you can pull the full history to get both.

A background conflict-resolver decides what supersedes what, and low-confidence contradictions get quarantined for review instead of silently overwriting.

So: latest by default, both on request, caller picks. "What to let go of" is salience + decay, and honestly the part I'm least done with.

[Gowry @gowry]

https://t.co/DcI2xiEJ3A

[Gowry @gowry]

Something I needed, not something I thought was cool. I switch between a bunch of AI tools all day and none of them share a memory, so I kept re-explaining myself to every one. Built afair for exactly that: one memory that’s mine, follows me across all of them. Open-sourced it last night.

[Gowry @gowry]

@rohanpaul_ai https://t.co/eZzGMRymrY

[Gowry @gowry]

Exactly the bet I built afair on, just one layer up: not a single agent’s task memory but your own durable memory across every tool you use. Append-only full-fidelity log, a compact interpreted view on top, dereference the exact event when you need it.

Been living on it myself for months, finally open-sourced it last night.

[Gowry @gowry]

@kennyhanson Here’s the link

https://t.co/eZzGMRymrY

[Gowry @gowry]

@kennyhanson Built afair for exactly this. Recall only surfaces what’s relevant to the thing in front of you, old stuff decays, and you can see and fix what shows up. So a fresh project doesn’t get tainted by some 11-day-old convo.

[Gowry @gowry]

@Angaisb_ Felt this so hard I built afair for it. Keeps everything, surfaces only what’s relevant when you ask, and you can see and fix what it kept. No filler in the window.

Happy about any PRs or any feedback

https://t.co/eZzGMRymrY

[Gowry @gowry]

@bcherny aren't we all kinda still in exploring mode overall? and your pre-PMF line kind of says it. 1+2+3, are the same people.

imo, roles feel like a later problem, once there's something to specialize around.

[Gowry @gowry]

Wo kommen eigentlich diese ganzen Klimaexperten hier auf X her, die vehement den menschengemachten Klimawandel abstreiten?

Sehe hier im Feed normalerweise tolle AI Posts, jetzt ganz viele dieser Deppenposts. Why?

[Gowry @gowry]

seeing these mcp tool calls (without me calling them) in any session and any agent feels like magic

[Gowry @gowry]

as far as I remember

https://t.co/8ytKosG4Uq

[Gowry @gowry]

1/ Your AI is a goldfish with a PhD.

Genius-level, and it forgets you after every single session. I got tired of reintroducing myself to a machine or its other friends. So I’m building afair.

[Gowry @gowry]

4/ It‘s close: this week I start inviting the first batch of test users.

Building it solo, in the open. If you bounce between AI tools all day and want in early, drop your email via web, that’s where the first invites are going.