CITATION.cff

cff-version: 1.2.0
message: "If CodexProLong, its PRO-LONG-inspired persistent-memory research architecture, or its results inform your work, please cite this repository and the PRO-LONG paper."
title: "CodexProLong: a persistent-memory research agent for long-horizon scientific problem solving"
type: software
authors:
  - family-names: Weatherhead
    given-names: James
repository-code: "https://github.com/JamesWeatherhead/CodexProLong"
license: MIT
date-released: 2026-08-15
version: 0.2.0
abstract: >-
  CodexProLong is a long-horizon research agent architecture that combines
  OpenAI Codex with a persistent filesystem research state inspired by
  PRO-LONG, web and scientific-literature search, executable program synthesis,
  iterative hypothesis testing, and verifier-guided research. The system
  maintains an accumulating research state across contexts, allowing later
  attempts to recover prior hypotheses, experiments, failures, programs,
  scores, and useful checkpoints rather than starting from scratch. In its
  first EinsteinArena campaign, CodexProLong used GPT-5.6 Sol through Codex and
  produced five domain-valid first-place constructions across 17 rankable
  mathematical problems in the frozen August 15, 2026 snapshot. This
  repository provides the public documentation, provenance, receipts, hashes,
  and selected evidence for that system and campaign; the full production
  research system is not included in the current public release tree.
references:
  - type: article
    title: "PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning"
    authors:
      - family-names: Fox
        given-names: Alexis
      - family-names: Wang
        given-names: Junlin
      - family-names: Rosu
        given-names: Paul
      - family-names: Dhingra
        given-names: Bhuwan
    year: 2026
    url: "https://arxiv.org/abs/2607.20064"