Kata Context

Status: v0.2.0 (Database + Storage Layer) — Storage foundation complete. Building in public.

What This Is

Kata Context is a standalone context policy engine for AI agents. It manages what goes in and out of the LLM context window — handling compaction, summarization, retrieval, and budget-aware windowing. Framework-agnostic: works with any agent system, or none.

Part of the Kata ecosystem, alongside Kata Orchestrator, Kata Agents, Kata Context.

Core Value

Policy, not storage. Given messages and a context budget, determine the optimal window to send to the model. The value isn't persisting conversations — it's intelligently managing finite context.

The Key Insight

Kata Context answers ONE question: "Given this conversation history and this token budget, what's the optimal context window to send?"

It doesn't care about:

• Which LLM you're using
• How your agent is structured
• What tools you have
• How you handle responses

┌─────────────────────────────────────┐
│         Your Agent Code             │
│  (LangChain, custom, Kata, etc.)    │
└─────────────────┬───────────────────┘
                  │ "What should I send?"
                  ▼
┌─────────────────────────────────────┐
│         Kata Context                │
│  - Stores full history              │
│  - Applies policy                   │
│  - Returns optimal window           │
└─────────────────┬───────────────────┘
                  │ Optimized context
                  ▼
┌─────────────────────────────────────┐
│         Any LLM                     │
│  (OpenAI, Anthropic, local, etc.)   │
└─────────────────────────────────────┘

Infrastructure, Not Framework

| Aspect | Framework (Letta, LangChain) | Infrastructure (Kata Context) | | ---------------- | ---------------------------- | ----------------------------- | | Relationship | You build inside it | You call it from outside | | Control | Framework controls lifecycle | You control lifecycle | | Opinions | Dictates agent patterns | No opinion on agent design | | Coupling | Tight — hard to leave | Loose — easy to swap | | Analogy | Rails | Postgres |

API

REST API for context management with token-budgeted retrieval:

# Create a context
curl -X POST https://your-app.vercel.app/api/v1/contexts \
  -H "Content-Type: application/json" \
  -d '{"metadata": {"agent": "my-agent"}}'

# Append messages
curl -X POST https://your-app.vercel.app/api/v1/contexts/{id}/messages \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "Hello", "tokenCount": 5}]}'

# Get optimal window for token budget
curl "https://your-app.vercel.app/api/v1/contexts/{id}/window?budget=8000"

SDKs for Python and TypeScript coming in v0.3.0.

Development

# Install dependencies
pnpm install

# Run tests (87 tests, 100% coverage)
pnpm test

# Lint and format
pnpm check

# Database commands
pnpm db:generate   # Generate migrations
pnpm db:migrate    # Apply migrations
pnpm db:studio     # Open Drizzle Studio

# Build
pnpm build

Stack: TypeScript 5.9, pnpm, Biome, Vitest, Drizzle ORM, PostgreSQL (Neon), pgvector, Vercel serverless

Roadmap

• [x] v0.1.0 — Core Setup (TypeScript, linting, testing, CI)
• [x] v0.2.0 — Database + Storage Layer (PostgreSQL, repository pattern, REST API)
• [ ] v0.3.0 — Policy engine (compaction, forking, semantic retrieval)
• [ ] v1.0 — Hosted API with multi-tenancy

License

Apache 2.0 (core engine and SDKs)