Dreamcatcher

The Problem

AI-assisted work generates enormous amounts of valuable context — conversations, decisions, half-formed ideas, research fragments — that immediately disappears the moment a session ends. Without a way to capture and organise this knowledge, every new session starts from zero.

Dreamcatcher is a canonical knowledge spine: it captures, organises, and contextualises AI interaction fragments, separating durable truth from ephemeral presentation and enabling semantic retrieval across distributed project contexts.

Architecture

Key Design Principles

• Truth vs presentation: Dreamcatcher separates the canonical fact (e.g. "API endpoint is /v2/users") from how it was expressed in context. This means updating a fact in one place propagates across all references.
• Automated topic clustering: Using sentence-transformers + HDBSCAN to automatically group related fragments into topic nodes without requiring manual tagging.
• Drift signals: The system surfaces knowledge that hasn't been validated recently or that contradicts newer entries — flagging potential drift before it causes downstream errors.
• Context assembly: For any given task, Dreamcatcher assembles a curated context bundle from the knowledge graph — reducing prompt tokens while maximising relevance.

Use in Practice

Dreamcatcher has been used to manage the knowledge base across multiple long-running AI development projects — capturing design decisions, architecture choices, and lessons learned in a way that's retrievable months later without manual organisation.

Access: the live demo is a frontend-only showcase on mock data — the cockpit, not the engine. The capture pipeline, knowledge store, and retrieval engine are private and available to discuss in a working session.