facebookresearch/context-language-models

This page introduces an external open-source repository. It is not an HDATF product.

What it does

The official repository for the Context Language Models paper, which describes language models that manage their own context by treating the context as a file the model may update without restriction. The authors report that building such models zero shot beats existing context management strategies across several tasks, including higher accuracy with fewer FLOPs on BrowseComp-Plus and on EdgeBench, and that the approach extends to multi agent settings where several agent contexts coexist as files.

How it helps ATF

Harness keeps the state and memory of a long AI run, and deciding what stays in context is the part that decides whether the run holds together. A paper that hands that decision to the model by writing context to a file is worth reading next to how Harness keeps memory now.

License

Custom license Custom or non-standard license. Read the license file before any reuse.

More in this category

• tigerless-labs/agent-memory
  A long-term memory runtime for AI agents. Plain Markdown files are the source of truth, a SQLite index is a deletable cache, local ranked retrieval returns file paths, and Claude Code and Codex share one store.
• kunchenguid/backpass
  A tool that reads local agent session transcripts from seven harnesses and proposes evidence-backed edits to AGENTS.md, CLAUDE.md and project skills under a token budget. Analysis never writes, and a person accepts or rejects each edit.
• Asymptote-Labs/agent-beacon
  A memory layer that captures coding-agent sessions across harnesses and turns workflows, corrections, and debugging patterns into reusable knowledge. Reviewed knowledge is available through MCP or Agent Skills.
• vectorize-io/hindsight
  An agent memory system focused on learning over time rather than only recalling conversations. Its documentation outlines retention, recall, reflection, observations, and memory banks.
• tamaratran/fast-jev-compaction
  A Claude Code plugin and npm library that uses Jev scoring to remove or truncate stale tool calls and results instead of summarizing conversation history. Retained content preserves its original wording.

Only repositories in the ranked Trendshift lists are included, and the lists are used only to find candidates. We do not copy their ranks. Descriptions, licenses and star counts come from each GitHub repository. The notes are our own reading. We have not tested these projects, and a place on a trending list does not prove quality.