PRO-LONG Deep Dive: Programmatic Memory Enables Long-Horizon Reasoning

· TopDigg Research Team · PRO-LONG / LLM Agent / Programmatic Memory / Long-Horizon Reasoning / ARC-AGI-3 / Context Management / Open Source / AI / Fable / Agent Memory

PRO-LONG is a minimal programmatic memory framework for LLM agents that enables long-horizon reasoning through single-file logs and code retrieval. This comprehensive analysis covers the project's architecture, design philosophy, practical tutorials, and core insights into agent memory systems.


1. Project Overview

1.1 What is PRO-LONG?

PRO-LONG is a minimal context management framework designed for long-horizon tasks. Its core idea is elegantly simple:

  1. Append every observation, action, and outcome to a single structured log.txt file
  2. Agent retrieves and reasons over this history programmatically (grep, Python)
  3. No subagents, no specialized retrieval mechanisms, system prompt is only ~30 lines

This is not another complex memory system. PRO-LONG's design philosophy is minimalism — achieving the most effective memory management with the least code.

1.2 Core Features

Feature Details
Single File Log All history stored in one log.txt file
Code Retrieval Agent uses grep, Python, and other tools to programmatically search history
Minimal Prompt System prompt is only ~30 lines, no complex instructions
Dual Backend Support Supports both OpenAI Codex and Claude Code backends
Docker Sandbox Executes in isolated container environments for security
ARC-AGI-3 Breakthrough Achieves 97.4% best@2 on ARC-AGI-3

1.3 Key Concepts

Programmatic Memory — Teaching Agents to "Look Things Up"

Traditional agent memory systems typically use two strategies:

  1. Context Injection: Putting all historical information directly into the prompt (causing token explosion)
  2. Vector Retrieval: Using embedding models to retrieve relevant history (adding complexity and latency)

PRO-LONG proposes a third strategy: programmatic memory. Agents can search and analyze history using tools like grep and Python scripts, just like programmers do.

Advantages of this approach:

  • Completeness: Preserves complete history without losing any information
  • Precision: Code retrieval is more precise than semantic retrieval
  • Interpretability: Agent's retrieval process is transparent and debuggable
  • Zero Overhead: No embedding models or vector databases needed

Single File Log — Simplest is Most Effective

PRO-LONG stores all information in a single log.txt file, including:

  • Initial board states
  • Board states after each action
  • Agent's analysis and reasoning
  • Action execution results

This design seems "naive" but is actually very clever:

  • No Information Loss: Complete preservation of all history
  • Simple and Reliable: No complex synchronization or indexing mechanisms
  • Efficient Retrieval: grep performs extremely well on large files

30-Line Prompt — Trusting Agent Capabilities

PRO-LONG's system prompt is only ~30 lines and does not include:

  • Complex reasoning instructions
  • Detailed strategy guidance
  • Specific task format requirements

It only tells the agent:

  1. What your goal is (solve puzzles)
  2. Where history is stored (log.txt)
  3. How to retrieve history (using code)
  4. How to output actions (write actions.json)

This minimal design reflects trust in agent capabilities — letting the agent decide how to retrieve and reason.


2. Design Philosophy

2.1 Minimalism — Less is More

PRO-LONG's core design philosophy is minimalism. While other memory systems continuously add complexity, PRO-LONG chooses the simplest solution:

  • One file stores all history
  • One prompt tells the agent how to use it
  • One set of tools lets the agent retrieve on its own

Advantages of this design:

  • Easy to Understand: Anyone can see how the system works
  • Easy to Debug: When problems occur, just check the log file
  • Easy to Extend: Adding new features only requires modifying the log format

2.2 Trusting Agents — Letting Code Speak

PRO-LONG doesn't try to "teach" agents how to reason. It trusts agent capabilities and only provides:

  • Access to history (file system)
  • Retrieval tools (grep, Python)
  • Output format (JSON)

Agents can:

  • Use any retrieval strategy
  • Write any analysis scripts
  • Adopt any reasoning method

This design reflects confidence in modern LLM coding abilities.

2.3 Programmatic Over Semantic — Precision Beats Ambiguity

Traditional memory systems use semantic retrieval (embedding similarity), but PRO-LONG chooses programmatic retrieval (grep, Python).

Reasons:

  • Exact Matching: grep can precisely find lines containing specific patterns
  • Structured Queries: Python can parse log formats and execute complex queries
  • Zero Latency: No embedding computation or vector search needed
  • Interpretable: Agent's retrieval process is completely transparent

3. Detailed Tutorial

3.1 Installation and Setup

Requirements

  • Python 3.12 (recommended)
  • Docker

Installation Steps

# Clone the repository
git clone git@github.com:alexisfox7/PRO-LONG.git
cd PRO-LONG

# Create virtual environment
python -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -e .

Build Docker Images

# Codex backend
docker build -t rgb-agent/codex-sandbox:latest docker/codex-sandbox
docker build -t rgb-openai-proxy docker/openai-proxy

# Claude Code backend
docker build -t rgb-agent/claude-sandbox:latest docker/claude-sandbox
docker build -t rgb-anthropic-proxy docker/anthropic-proxy

Configure Environment Variables

Create .env file:

ARC_API_KEY=...
ANTHROPIC_API_KEY=...   # claude-code backend
OPENAI_API_KEY=...      # codex backend

3.2 Basic Usage

Running Evaluation

# Run all games with Codex backend
prolong-swarm --suite all -m gpt-5.5 --max-actions 500

# Run all games with Claude Code backend
prolong-swarm --suite all --backend claude-code -m claude-opus-4-6

# Run specific games
prolong-swarm --game ls20,ft09 -m gpt-5.5

Key Parameters

Parameter Default Description
--backend codex Backend: codex or claude-code
--suite Game suite: ls20, vc33, ft09, or all
--game Comma-separated game names or IDs
--max-actions 500 Maximum actions per game
--model, -m claude-opus-4-6 Base model
--effort high Effort level (claude-code backend)
--reasoning-effort none Reasoning effort (codex backend)
--operation-mode online online / offline / normal

3.3 Memory Conditions

Agent access to game history is controlled by --log-window and --workspace:

Condition Flags Available History
prolong (default) Full game log
lw25 --log-window 25 Last 25 action sections of the log
no-log (in-prompt) --log-window -1 No log file; current board added to prompt
stateless --workspace stateless Full log, but workspace wiped each call

3.4 Understanding the System Prompt

PRO-LONG's system prompt is very concise, with core content:

SYSTEM_PROMPT = """
You are a coding agent playing a grid-based puzzle game by writing Python action plans.

Your primary objective is to solve all levels in the game. Your secondary objective is to minimize total cumulative actions used.

`/workspace/logs.txt` is the game log: action headers, tool calls, board states, and your own prior analyses. Parse it **programmatically**, as reading full 64x64 board states from prompt can introduce precision errors.

**Tools**: Read, Write, Edit, Bash, Grep, Glob.

**Workspace**: `/workspace/` persists across calls. `actions.json` is cleared each call; other files accumulate.

**Response format**: a strategic briefing, then
[PLAN]
<2-3 sentence action plan>

**Write `/workspace/actions.json`** with a JSON object `{"actions": ["ACTION6(30,40)", "ACTION1", "RESET"]}` — a list of 1–{action_cap} actions to execute in order.
"""

Key points of this prompt:

  1. Clear Goal: Solve puzzles + minimize action count
  2. 指定记忆位置/workspace/logs.txt
  3. 指定检索方式:程序化(grep、Python)
  4. 指定输出格式actions.json

3.5 Action System

PRO-LONG supports the following actions:

Action Description
ACTION1 Up
ACTION2 Down
ACTION3 Left
ACTION4 Right
ACTION5 Spacebar / interact
ACTION6(x,y) Click at column x (0-63), row y (0-63)
ACTION7 Undo
RESET Reset level (actions still count)

3.6 Output Results

Evaluation results are written to evaluation_results/ directory. The scorecards/ directory contains official online scorecards.


4. Core Architecture Deep Dive

4.1 Project Structure

prolong_agent/
├── agent/
│   ├── base.py               # Base architecture
│   ├── codex_agent.py        # Codex CLI backend
│   ├── claude_code_agent.py  # Claude Code backend
│   ├── swarm.py              # CLI entry point
│   ├── action_queue.py       # Action execution
│   ├── game_state.py         # Board/log formatting
│   └── prompts.py            # Prompt templates (~30 lines)
├── environment/
│   ├── arcagi3.py            # ARC-AGI-3 API wrapper
│   ├── runner.py             # Per-game loop
│   └── config.py
├── metrics/
└── utils/

4.2 Core Components

Agent Base Architecture

class BaseAgent:
    """Base agent class defining standard interface"""
    
    def __init__(self, model: str, workspace: str):
        self.model = model
        self.workspace = workspace
        self.log_path = f"{workspace}/logs.txt"
    
    def act(self, observation: dict) -> list[str]:
        """Return action list based on observation"""
        # 1. Append observation to log
        # 2. Read log
        # 3. Use model to generate actions
        # 4. Write actions.json
        pass

Log Format

[INITIAL BOARD STATE]
<64x64 board state>

[ACTION1]
Tool call: bash("python3 -c '...'")

[POST-ACTION BOARD STATE]
<updated board state>

[ACTION2]
Tool call: grep("pattern", "/workspace/logs.txt")
...

Action Execution

class ActionQueue:
    """Action queue, executing actions in order"""
    
    def execute(self, actions: list[str]) -> dict:
        results = []
        for action in actions:
            result = self._run_action(action)
            results.append(result)
        return {"results": results, "total": len(results)}

4.3 Retrieval Mechanism

PRO-LONG's retrieval relies entirely on agent code capabilities:

# Retrieval methods available to agents

# 1. grep search for specific patterns
grep -n "INITIAL BOARD STATE" /workspace/logs.txt

# 2. Python log parsing
python3 -c "
import re
with open('/workspace/logs.txt') as f:
    content = f.read()
boards = re.findall(r'\[POST-ACTION BOARD STATE\](.*?)\[', content, re.DOTALL)
print(f'Found {len(boards)} board states')
"

# 3. Statistical analysis
python3 -c "
with open('/workspace/logs.txt') as f:
    lines = f.readlines()
actions = [l for l in lines if l.startswith('[ACTION')]
print(f'Total actions: {len(actions)}')
"

4.4 Performance Data

According to the paper and official evaluation:

Metric Data
ARC-AGI-3 best@2 97.4% (Fable 5)
Average Improvement +18.0 percentage points over base agent
Token Efficiency 4.2-5.8x fewer than specialized frameworks
Total Cost $1,750 (25 Fable 5 runs)
Highest pass@1 76.1%

5. Summary of Insights

5.1 Why PRO-LONG Matters

PRO-LONG represents an important paradigm shift in agent memory systems. While other systems continuously add complexity, PRO-LONG proves the power of minimalism.

Three Core Insights:

  1. Programmatic Memory Beats Semantic Retrieval: Letting agents search history with code is more precise and efficient than embedding retrieval
  2. Single File Log is Enough: One log.txt file can store all needed information
  3. Trust Agent Capabilities: A 30-line prompt is sufficient for agents to autonomously complete complex tasks

5.2 Comparison with Other Tools

Feature PRO-LONG LangChain Memory AutoGPT BabyAGI
Memory Method Single file log Vector database Multiple files Task queue
Retrieval Method Code (grep/Python) Semantic search File reading Priority sorting
Prompt Length ~30 lines Complex Complex Medium
Token Efficiency Extremely high Medium Low Medium
ARC-AGI-3 97.4% Not tested Not tested Not tested
Open Source

5.3 Use Cases

Best For:

  • Agent tasks requiring long-term memory
  • Historical queries requiring precise retrieval
  • Complex reasoning and planning tasks
  • Cost-sensitive application scenarios

Less Suitable For:

  • Simple single-turn conversations
  • Tasks not requiring historical memory
  • Non-coding agents (requires coding capabilities)

5.4 Design Philosophy Summary

PRO-LONG's design philosophy can be summarized as:

  1. Minimalism: Least code, most effective memory
  2. Trust Agents: Let agents decide how to retrieve and reason
  3. Programmatic Over Semantic: Exact matching beats fuzzy similarity
  4. Complete Preservation: Don't lose any historical information
  5. Zero Extra Overhead: No embedding models or vector databases needed

6. Roadmap

Based on project trends and evolution in agent memory systems:

Short-Term (3-6 months)

  • Support more LLM backends
  • Improve log format and retrieval efficiency
  • Add more evaluation benchmarks

Medium-Term (6-12 months)

  • Multi-agent collaborative memory
  • Incremental log compression
  • Cross-session memory persistence

Long-Term (1-2 years)

  • Autonomous memory management agents
  • Cross-organization memory sharing
  • General long-horizon reasoning framework

7. Conclusion

PRO-LONG is a groundbreaking agent memory framework that achieves breakthrough performance through minimal design. Single file log, code retrieval, 30-line prompt — these seemingly "naive" designs achieved 97.4% accuracy on ARC-AGI-3.

Core Value:

  • Minimalism: Least code, most effective memory
  • Programmatic Retrieval: Precise, efficient, interpretable
  • Complete Preservation: Don't lose any historical information
  • Zero Extra Overhead: No embedding models needed

Why Choose PRO-LONG?

  • Open and transparent (MIT License)
  • Minimal design, easy to understand and debug
  • Code retrieval, precise and efficient
  • Breakthrough performance validated on ARC-AGI-3

Get Started:

# Clone the repository
git clone git@github.com:alexisfox7/PRO-LONG.git
cd PRO-LONG

# Install
python -m venv .venv
source .venv/bin/activate
pip install -e .

# Run evaluation
prolong-swarm --suite all -m gpt-5.5 --max-actions 500

Disclaimer: This article is based on PRO-LONG's public documentation, paper, and technical analysis, aiming to provide comprehensive technical insights and practical guidance. Paper citation: arXiv:2607.20064.