Programmatic memory for long-horizon LLM agents

rzk1 pts0 comments

GitHub - alexisfox7/PRO-LONG: Programmatic memory for long-horizon LLM agents: the harness appends everything to one log, and the agent searches it with code. 97.4% on ARC-AGI-3 (arXiv:2607.20064) · GitHub

/" data-turbo-transient="true" />

Skip to content

Search/

Sign in<br>Sign upAppearance settings

You signed in with another tab or window. Reload to refresh your session.<br>You signed out in another tab or window. Reload to refresh your session.<br>You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

{{ message }}

alexisfox7

PRO-LONG

Public

Notifications<br>You must be signed in to change notification settings

Fork<br>40

Star<br>399

main

BranchesTags

Go to file

CodeOpen more actions menu

Folders and files<br>NameNameLast commit message<br>Last commit date<br>Latest commit

History<br>42 Commits<br>42 Commits

assets

assets

docker

docker

prolong_agent

prolong_agent

release_logs/fable5

release_logs/fable5

scorecards

scorecards

.gitignore

.gitignore

.python-version

.python-version

CITATION.cff

CITATION.cff

LICENSE

LICENSE

README.md

README.md

pyproject.toml

pyproject.toml

uv.lock

uv.lock

View all files

Repository files navigation

PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning

PRO-LONG is a minimal memory addition for LLM agents on long-horizon tasks. The harness appends every observation, action, and outcome to a single structured log.txt, and the agent retrieves and reasons over it programmatically (grep, Python). There are no subagents or specialized retrieval mechanisms, and the system prompt is about 30 lines.

On the full ARC-AGI-3 public game set, PRO-LONG improves over the same coding agents without the log by 18 percentage points on average, matches or exceeds specialized harnesses at 4.2–5.8x fewer billed tokens, and reaches 97.4% best@2 with Fable 5 at a total cost of $1,750.

Paper: arxiv.org/abs/2607.20064

Setup

Requires Python (3.12 recommended) and Docker.

git clone git@github.com:alexisfox7/PRO-LONG.git<br>cd PRO-LONG<br>python -m venv .venv<br>source .venv/bin/activate<br>pip install -e .

# codex backend<br>docker build -t prolong-agent/codex-sandbox:latest docker/codex-sandbox<br>docker build -t prolong-openai-proxy docker/openai-proxy

# claude-code backend<br>docker build -t prolong-agent/claude-sandbox:latest docker/claude-sandbox<br>docker build -t prolong-anthropic-proxy docker/anthropic-proxy

Create a .env file:

ARC_API_KEY=...<br>CODEX_API_KEY=... # codex backend<br>CLAUDE_CODE_OAUTH_TOKEN=... # claude-code backend (default)<br>ANTHROPIC_API_KEY=... # claude-code backend with --api-key

The agent container only mounts the game workspace and, by default, has no network access except a proxy to the model API.

Usage

prolong-swarm --suite all -m gpt-5.5 --max-actions 500<br>prolong-swarm --suite all --backend claude-code -m claude-opus-4-6<br>prolong-swarm --game ls20,ft09 -m gpt-5.5

Results are written to evaluation_results/.

Key flags

Flag<br>Default<br>Description

--backend<br>codex<br>codex (OpenAI Codex CLI) or claude-code (Claude Code CLI)

--suite<br>Game suite (all only)

--game<br>Comma-separated individual game names or full IDs

--max-actions<br>500<br>Max actions per game

--model, -m<br>Backend-specific<br>gpt-5.5 for Codex; claude-opus-4-6 for Claude Code

--effort<br>high<br>Effort level (claude-code backend)

--reasoning-effort<br>none<br>Reasoning effort (codex backend)

--operation-mode<br>online<br>online / offline / normal

Memory conditions

The agent's access to game history is controlled by --log-window and --workspace. These are the ablation conditions from the paper:

Condition<br>Flags<br>History available

prolong<br>(default)<br>Full game log

lw25<br>--log-window 25<br>Last 25 action sections of the log

no-log (in-prompt)<br>--log-window -1<br>No log file; the current board is added to the prompt

stateless<br>--workspace stateless<br>Full log, but the workspace is wiped each call

Scorecards & logs

scorecards/ contains the official online scorecards, including all 25 Fable 5 runs from the paper (fable_online_scorecards.txt); each can be verified on arcprize.org. release_logs/ contains logs for the Fable 5 online runs: game logs, agent transcripts, and workspaces. Logs for the remaining ablations will be added.

Architecture

prolong_agent/<br>├── agent/<br>│ ├── base.py # base architecture<br>│ ├── codex_agent.py # Codex CLI backend<br>│ ├── claude_code_agent.py # Claude Code backend<br>│ ├── swarm.py # CLI entry point<br>│ ├── action_queue.py # action execution<br>│ ├── game_state.py # board/log formatting<br>│ └── prompts.py # prompts (~30 lines)<br>├── environment/<br>│ ├── arcagi3.py # ARC-AGI-3 API wrapper<br>│ ├── runner.py # per-game loop<br>│ └── config.py<br>├── metrics/<br>└── utils/

This repo was formerly the Read-Grep-Bash (RGB) Agent, see our original blog post on the ARC-AGI-3 preview games.

Citation

@misc{fox2026prolong,<br>title={PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning},<br>author={Fox, Alexis and Wang, Junlin and Rosu, Paul and Dhingra,...

long claude backend docker game code

Related Articles