Most AI applications treat conversations as fixed systems. You can choose a model, but the application decides how prompts are built, how context is managed, and how tools are used.

Looma takes a different approach.

The platform provides the infrastructure — models, storage, and tools — while strategies define how conversations work.

---

Conversations Are Context Systems

Large language models are fundamentally context-limited systems. Every response depends on what information is placed into the prompt. This means every AI application must constantly answer questions like:

• What conversation history should be included?
• What information should be removed to fit token limits?
• When should knowledge be retrieved?
• When should the system call tools?

Most applications hide this logic inside the product.

Looma exposes it.

---

Programmable Chat Flow

In Looma, the behavior of a conversation is defined by a strategy.

A strategy is a small TypeScript module that controls how the system interacts with the model.

Strategies define the chat flow of a conversation:

• how prompts are constructed
• which context should be included
• how history is selected
• when tools are used
• when memories are written or retrieved
• and everything...

Because strategies run inside an isolated engine, they can extend the system without modifying the core application.

---

Local-First by Design

Looma is designed around a local-first architecture.

Conversations, files, and indexes are stored locally using a single local file. This keeps the system fast and private while still allowing cloud models when needed.

The platform handles the heavy infrastructure work:

• model provider integration
• token budgeting
• tool execution
• file ingestion
• retrieval-augmented generation

Strategies only focus on defining conversation behavior.

---

System Architecture

Looma is organized into three layers.

Host Platform

Handles the UI, database, model APIs, and system lifecycle.

Strategy Engine

Runs programmable strategies inside isolated worker threads.

Memory & Ingest System

Manages document ingestion, vector indexing, and semantic search.