Skip to main content

The problem with building AI features

Most teams integrate LLMs directly — glue code, prompt templates, manual context injection. It works for demos. It breaks in production.
  • No memory — every conversation starts from zero
  • No user isolation — concurrent users share state or fight over it
  • No autonomy — agents only run when users ask
  • Ops overhead — you own the infra, the retries, the scaling
Kairn is the layer between your product and the LLM. It handles the hard parts so you ship the agent, not the scaffolding.

What Kairn gives you

Per-user memory

Every user gets a persistent key-value store. The agent reads it at invocation time, writes to it during the session. Memory persists across conversations — no re-seeding required.

Soul-based identities

Define agent identities once. One soul serves all your users with zero shared state. Change the soul’s instructions and every user gets the update instantly.

Streaming out of the box

kairn.stream() yields typed events — tokens, tool calls, memory writes, done. Wire directly into your Next.js route and stream to the browser.

Mandates — scheduled agents

Agents that run on their own. Cron-triggered, user-scoped, memory-aware. Weekly check-ins, daily reports, proactive nudges — without user interaction.

How it works

1

Create a soul

Define the agent’s identity, instructions, and tool access. One soul powers all your users.
2

Seed memory from your DB

Before each conversation, write user context from your database into the agent’s memory store. Name, plan, recent actions — whatever the agent needs to personalise.
3

Invoke or stream

Call kairn.invoke() or kairn.stream() from your API route. The agent picks up context, runs, and writes what it learned back to memory.
4

Read back what the agent wrote

After each turn, memoryDelta tells you exactly what the agent updated. Sync it back to your own DB or surface it directly in your UI.

Pick your path

Quickstart

From zero to your first agent invocation in under 5 minutes.

Core concepts

Souls, memory, sessions, mandates — the mental model explained.

SDK reference

Full API for every method, type, and error.

Use cases

Real product scenarios with working code you can adapt.