Kin remembers what matters

Most assistants forget you between sessions. You re-explain that you're vegetarian. You re-mention that your sister Maya lives in Berlin. You re-introduce your team every time you ask about scheduling.

Kin has long-term memory. Tell it something once and it surfaces it when relevant.

How it works

After every assistant turn, Kin runs a quick background pass that asks: "Was anything in this exchange worth remembering for next time?" The answer is usually no (the model is conservative on purpose). When the answer is yes, the model extracts 1–3 short facts and embeds each into your private vector store.

Example facts that get extracted:

• "User's sister Maya lives in Berlin and studies architecture." (relationship)
• "User prefers concise replies." (preference)
• "User is launching a SaaS called PersonaliAI." (fact)
• "User's Expenses 2026 sheet has columns Date, Category, Description, Amount." (fact)

Example facts that don't get extracted (correctly):

• "User asked about today's weather." (transient)
• "User said 'thanks'." (no signal)
• "User has 47 unread emails." (tool result, not a stable fact)

How retrieval works

On every new message, Kin embeds your question and looks up the top 5 most-similar memories. If they're above the similarity threshold (0.55), they get injected into the system prompt as: "Things you remember about this user: …"

The agent now has context it didn't have last conversation.

You (last week): I'm vegetarian and I dislike sushi.

Kin: Got it.

---

You (this week): Recommend a restaurant for tomorrow's team dinner near the office.

Kin: Given you're vegetarian (and not a fan of sushi), I'd suggest [restaurant] — they have a strong tasting menu with plant-forward options.

You didn't have to repeat anything.

Active recall

The agent can also explicitly search its memory mid-turn via the recall_memory tool:

You: What did I tell you about my brother's wedding?

Kin (calls recall_memory with "brother wedding") → finds: "User's brother is getting married on June 14 in Lisbon."

You're in control

Visit /dashboard/memory to see exactly what Kin remembers. Each entry shows the kind (preference / fact / event / relationship / goal / habit), the content, and when it was added. You can:

• Add a memory manually ("teach Kin something")
• Delete an individual memory
• Forget everything (nuclear reset)
• Turn off long-term memory entirely from Settings — Kin reverts to short-term context only

Privacy guarantees

Memories live in your row in our database. They're isolated per user via row-level security. They're never sent to advertisers, never used to train models. The embeddings model (text-embedding-004) doesn't retain inputs.

Cost

Each turn adds roughly $0.0006 — about an eleventh of a cent. Worth it for an assistant that doesn't lose its mind every conversation.