Introduction to Recurse
Recurse is infrastructure for sense-making, not just retrieval. It structures knowledge for both humans and AI applications to enable exploration and deep understanding.
Most AI memory systems work by looking at query similarity – you ask a question, get back stuff that looks similar to your question, and that's it. This works if you already know what you're looking for. But it systematically prevents discovery. You can't find connections you didn't know existed, can't stumble onto context from unexpected places, can't explore ideas that branch away from your original query.
Recurse is built on different principles: structure over similarity, relationships over rankings, evolution over static storage. We are building infrastructure for sense-making, not just retrieval. For exploration, not just question-answering. For building understanding, not just finding facts.
How Recurse Works
Instead of similar chunks (RAG) or entities (GraphRAG), Recurse extracts semantic structure – understanding what sources argue, what supports those arguments, how different pieces relate – and maps out how these elements connect.
This structure lets information reference other information across your entire knowledge base. An argument in one paper connects to supporting data from three others. A technique described in one source links to examples of its use elsewhere. Your knowledge becomes a web of relationships, not a collection of isolated text chunks.
When you (or your AI agent) query Recurse, you don't just get back similar text – you can trace connections, follow supporting evidence, explore how ideas relate. Discover connections that weren't even explicit in the original sources.
Core Mechanisms
The system works through several interconnected mechanisms:
Frames
Typed semantic units with named slots that capture meaning and relationships.
The fundamental building blocks that enable navigation by structure. Frames reference each other and nest recursively to form knowledge graphs.
Adaptive Schemas
Schema discovery without predefined ontologies.
The system learns frame patterns from your content automatically. Creates emergent structure that adapts as you add sources across different domains.
Temporal Versioning
Evolution tracking with complete history preservation.
New information updates previous understanding while maintaining timestamps, diffs, and explanations. Creates living memory that stays current without losing context.
Source Subscriptions
Continuous, automatic ingestion from individual feeds.
Monitor URLs and RSS for ongoing updates. Temporal versioning maintains both current understanding and complete evolution history as sources change.
Context Streams
Domain expertise sharing based on trust and authority.
Experts share subgraphs of their knowledge bases—sources, documents, and semantic connections. Subscribe to streams from people you trust for queryable expertise.
RAGE
The processing engine that transforms unstructured sources into structured knowledge graphs.
Handles source ingestion, frame extraction, and relationship mapping. Creates evolving structures that navigate meaning rather than just matching keywords.
How These Work Together
Frame extraction provides structured semantic units. Adaptive schemas enable the structure to emerge from your content. Temporal versioning maintains knowledge currency while preserving evolution history. Source subscriptions keep information flowing automatically. Context Streams package expertise for sharing. RAGE integrates all these mechanisms into a coherent system.
The architectural consequence: Recurse treats knowledge as living structures rather than static text. This enables retrieval systems that navigate relationships rather than just matching keywords, accumulate genuine understanding rather than aggregating text, and support actual inquiry rather than just answering questions.