Skip to main content
Ryan Orban

Ryan Orban

Subject
5 entries

Context Window

Bookmarks

  1. Context Mode: MCP server for context window preservation

    Context Mode is an MCP server that intercepts large tool outputs and keeps them out of the context window — 98% token reduction, with BM25-indexed session history for continuity after compression. Addresses the 40%-context-consumed-in-30-minutes problem.

  2. Mastra Observational Memory

    Mastra's Observational Memory keeps agent context windows small by summarizing and compressing observations rather than keeping raw conversation history. A practical approach to the long-term vs. working memory tradeoff in AI agents.

  3. Context Lens: see what's filling your context window

    Context Lens is a local LLM API proxy that shows what's filling your context window — system prompts vs. tool definitions vs. conversation history vs. tool results. Answers the question every developer asks: why is this session so expensive?

  4. HyperContext: spatial context visualization for AI sessions

    HyperContext visualizes AI session state as a spatial map — see what the model actually knows, how the context window is structured, and where attention is distributed. Makes the invisible context window visible.

  5. Landmark Attention: Random-Access Infinite Context

    Landmark Attention paper extending LLaMA 7B to 32k token context by using landmark tokens to represent blocks of input, enabling attention-based retrieval of relevant blocks rather than attending over all tokens. Achieves near-GPT-4 context length through architectural change rather than longer pretraining.

All bookmarks