AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents

2026-07-02 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

AgenticSTS introduces a novel bounded-memory testbed designed for long-horizon LLM agents, addressing the challenge of managing context in extended decision-making sequences. Unlike traditional methods that append all past observations and tool calls, AgenticSTS employs a "bounded contract" where each decision is made from a fresh user message, assembled via typed retrieval. This approach ensures prompts remain constant in size regardless of run length, allowing for isolated ablation of memory components. The testbed is instantiated in Slay the Spire 2, a complex deck-building game requiring hundreds of strategic decisions. While frontier LLMs currently achieve zero wins at the lowest difficulty (compared to a 16% human win rate), an ablation study within AgenticSTS demonstrated that enabling strategic skills improved win rates from 3/10 to 6/10 games. The authors release a reproducible testbed comprising 298 completed trajectories, frozen memory/skill snapshots, and analysis scripts.

Key takeaway

For Machine Learning Engineers designing long-horizon LLM agents, traditional memory approaches that append all past context can lead to jumbled prompts and difficulty isolating component effects. You should explore implementing a bounded memory contract, where each decision uses fresh, retrieved context. This methodology, exemplified by AgenticSTS, allows for clearer analysis of individual memory layers and offers a scalable path for agents tackling complex, multi-step tasks like strategic games.

Key insights

A bounded memory contract for LLM agents, using typed retrieval for fresh context, enables isolated memory component study.

Principles

• Memory as a "bounded contract" prevents context bloat.
• Isolate memory components for clearer impact analysis.

Method

Decisions are made from fresh user messages, assembled by typed retrieval, without appending raw cross-decision transcripts to the prompt.

In practice

• Use Slay the Spire 2 as a complex, long-horizon LLM agent testbed.
• Ablate specific memory layers to quantify their impact on agent performance.

Topics

• LLM Agents
• Bounded Memory
• Long-Horizon Tasks
• AgenticSTS
• Memory Management
• Slay the Spire 2

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — LLM agent memory should be actively reconstructed and dynamically adapted to reasoning context, not statically retrieved.
• LongMINT: Evaluating Memory under Multi-Target Interference in Long-Horizon Agent Systems — Current memory-augmented agents struggle with interference and multi-target reasoning in long, dynamic contexts.
• SWE-MeM: Learning Adaptive Memory Management for Long-Horizon Coding Agents — Long-horizon coding agents can achieve superior performance and efficiency through adaptive, proactive memory management.
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective — EvoMemBench systematically evaluates LLM agent memory across scope and content, revealing current limitations and specialized strengths.
• Your Agent Forgot Everything Again. Here’s Why That’s a Design Problem. — LLM agents lack episodic memory, hindering their ability to learn from past interactions and maintain context.
• AriadneMem: Threading the Maze of Lifelong Memory for LLM Agents — AriadneMem is a novel structured memory system designed for long-horizon LLM agents, specifically addressing challenges of "disconnected evidence" and "state updates" within fixed context budgets.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.