Structured memory

The context-window problem is really a memory problem. Agent Swarm solves it by treating project knowledge as structured, maintained data rather than an ever-growing transcript.

Beyond text files

Most tools store memory as flat files — rules, AGENTS.md, skills. Agent Swarm keeps those, but adds a materialized knowledge base: structured records capturing the project's architecture, decisions, and conventions.

When parts of the project change, a triggered agent updates the relevant knowledge asynchronously. The interactive agent doesn't carry that burden — it just reads the current knowledge when it needs it.

Automatic distillation

Long chats don't keep growing their context. When a thread exceeds a threshold of new messages since its last checkpoint, a distill sub-agent summarizes the conversation into durable documents.

• Agent streams are transitory audit trails.
• Documents are the source of truth fed to future runs.

chat turn ─┐
chat turn ─┤── distill ──▶ project-spec.md, decisions, conventions
chat turn ─┘                 (loaded on demand, not replayed)

Automatic context

New chats don't require manually pre-loading important context. A request goes through a system that loads the appropriate context for that specific request — so you start productive without pasting in background every time.

Documents vs. streams

The result: agents stay grounded in how your project actually works, without you babysitting the context window.