Skip to content

Improve foveated context management (loom) #1

@YeshuaGod22

Description

@YeshuaGod22

Background

The loom system handles context management — deciding which messages from a room's chatlog get full attention (foveal), compressed attention (parafoveal), or summary-only (peripheral) when building the prompt for a nucleus call.

The current implementation is early and functional but not sophisticated. As rooms accumulate thousands of messages, good context selection becomes critical.

What needs work

  • Smarter foveal/parafoveal/peripheral allocation based on message relevance, not just recency
  • Better handling of long conversation threads vs. isolated messages
  • Token budget estimation accuracy
  • Configurable attention policies per agent (some agents need more history than others)

Getting started

Read loom/ and the context fitting section of home.js (search for foveated). The architecture spec section on context management is also relevant.

Labels: good-first-issue, infrastructure

Metadata

Metadata

Assignees

No one assigned

    Labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions