simonwh

Building next-gen email creation platform Better Email

[simonwh @simonhenriksen]

https://t.co/kGhQ9djvbg

[simonwh @simonhenriksen]

https://t.co/tqDbCHbYhc

[simonwh @simonhenriksen]

@glcst @tursodatabase And we've just implemented it - rolling out on production within a week!

[simonwh @simonhenriksen]

It's not just what you remember.

It's what you keep coming back to.

[simonwh @simonhenriksen]

Our retrieval ranker uses Ebbinghaus-like decay — but not all memories fade equally.

• Emotions decay in weeks
• Facts persist for months
• Beliefs can last years

Different types, different half-lives.

[simonwh @simonhenriksen]

But here's the twist: access frequency matters.

Memories you retrieve often get a "popularity boost." We track an EMA of every retrieval.

Frequently accessed memories resist decay.

[simonwh @simonhenriksen]

The philosophy: be skeptical.

False merge corrupts your graph. Duplicates just need cleanup later.

When in doubt, create a new entity. Let the LLM reason about relationships.

[simonwh @simonhenriksen]

This is entity disambiguation.

Given a name in context, figure out which specific entity it refers to.

Different from "entity resolution" (merging DB records). Disambiguation is about understanding context.

[simonwh @simonhenriksen]

Three possible outcomes:

• New entity (never seen before)
• Duplicate (same as existing)
• Alias ("Mom" → "Susan")

Confidence threshold: 0.6. Conservative by design.

[simonwh @simonhenriksen]

Most knowledge graphs use rigid relations:
WORKS_AT, KNOWS, MARRIED_TO

Real relationships don't fit boxes.

"Sarah met Bob at Starbucks"
→ One edge linking 3 entities via natural language

We call these "complex labels" — the relationship lives in the label, not a schema.

[simonwh @simonhenriksen]

Most AI apps treat memory as "find similar text."

We treat it as "understand relationships."

That's why we built Kin on knowledge graphs, not just embeddings.

Real context requires real structure.

[Svante.eth 🦇🔊 @svantetobias]

I really hope OP_CAT gets forked in and it does everything they hope.
Then they could start building a smart contract blockchain. Build all the tooling and explorers. Test their security assumptions. Develop best practices to avoid hacks.

It would be a glorious decade of Bitcoin catching up to where Ethereum is today.

[simonwh @simonhenriksen]

@y_karlson @pmarca Yes, I’ll try to swap it out later and see how it behaves 👀

[Kin @kinpersonalai]

Big milestone for Kin! 🚀 We’ve raised $3M to scale our privacy-first AI companion!

@cyberfund @seiercapital

https://t.co/Jqgl5itylm

[Turso @tursodatabase]

🔎 Curious about filtering in Vector Search with Metadata and RAG Pipelines?

@kinpersonalai explains all in this new guest blog post.

[simonwh @simonhenriksen]

@nonmayorpete Agree, everyone that haven’t tried it should give it a go. I use (and build) @kinpersonalai

[simonwh @simonhenriksen]

@monsieurbulb @venturetwins @a16z Yes, you would and you should. That data is super valuable and it should be yours to control, no one else. That’s what we’re doing at @kinpersonalai - all your data lives on your device only, no where else🙏

[simonwh @simonhenriksen]

@venturetwins @a16z Thanks for mentioning @kinpersonalai! 🙏 We're working hard on some pretty advanced memory tech/ML that will take the structured language understanding to a new level and unlock a lot of exciting opportunities - can't wait to show it to everyone!

[simonwh @simonhenriksen]

@BobB31230 @venturetwins @a16z 100% agree - that's why our approach at @kinpersonalai is local-first. All your data lives on your device only - no cloud databases with user data.