World's Fair 2026
RLM: Recursive Language Models for Large Codebases - Shashi, Superagentic AI
Overview
This talk introduces Recursive Language Models (RLMs) as a method to address the context window limitations of AI coding agents when working with large codebases. The core thesis is to externalize context management into a programmable execution environment, allowing models to operate on the repository as data, curate relevant chunks, and feed them into the main context window. This approach can also serve as a memory layer for coding agents.
Who should watch
- AI Engineers working with large codebases or monorepos.
- Developers experiencing performance degradation with coding agents as context grows.
- Builders looking for new patterns to manage context and memory for AI agents.
- Those interested in open-source implementations of advanced AI agent concepts.
Key takeaways
- Traditional coding agents struggle with large codebases due to context window limits, leading to performance degradation.
- Existing solutions include search tools (like grep), semantic search, context compression, and memory solutions.
- RLM's core idea is to create a separate, programmable environment where the model can write code to inspect, slice, and compute relevant context chunks.
- This process is analogous to a senior engineer navigating a large codebase by making notes, understanding structure, and consulting experts.
- The recursive aspect involves the model querying other models or systems for information, similar to asking a specialist.
- Codebases are structured data with directories, tests, and configurations, requiring models to reason over this structure, not just text.
- RLM code is an open-source reference implementation demonstrating RLM concepts, allowing integration with local or cloud models and observability frameworks.
- RLM concepts are being adopted in proprietary systems for managed agents, dynamic workflows, and software factories.
Notable quotes
*The core thesis of the RLM is you need to externalize the context management into a programmable execution environment.*
*So basically rather than putting everything into the models context create a separate dedicated environment give them a coding agent or ripple and then model write the code to curate the context that can be used into the main.*
Unofficial community note. Prefer the recording for nuance.