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

Hugging Face Daily Papers Papers
long-horizon llm-agents memory testbed bounded-memory typed-retrieval slay-the-spire

Summary

AgenticSTS introduces a bounded-memory testbed for long-horizon LLM agents using typed retrieval, with results on Slay the Spire 2 showing improved performance.

Memory for a long-horizon LLM agent is a contract about what each future decision is allowed to see. The simplest contract appends past observations, tool calls, and reflections to every prompt, which makes prior context easy to access but also turns it into a jumbled mixture in which the effect of any single memory component is hard to isolate. We introduce and instrument an alternative bounded contract: every decision is made from a fresh user message assembled by typed retrieval, with no raw cross-decision transcript appended. The prompt thus stays bounded across runs of any length, and any single layer can be ablated in isolation. We instantiate the contract in Slay the Spire 2, a closed-rule stochastic deck-building game whose runs require hundreds of tactical and strategic decisions. A public online benchmark of frontier LLMs on the same game reports zero wins at the lowest difficulty across five configurations, and the developer-reported human win rate at the same difficulty is 16%; the task is hard but not saturated. Within our harness, a fixed-A0 ablation shows the largest observed difference when triggered strategic skills are enabled: the no-store baseline wins 3/10 games and adding the skill layer 6/10. At this sample size the comparison is directional rather than statistically decisive (Fisher exact p\approx0.37); a cross-backbone probe and public accumulating-context baselines are reported as operational comparisons rather than controlled tests of the contract variable itself. We release a reproducible testbed: 298 completed trajectories with condition tags, frozen memory/skill snapshots, prompt records, and analysis scripts -- an agent design and a validated, reusable methodology for studying how explicit memory layers shape long-horizon LLM-agent decisions.
Original Article
View Cached Full Text

Cached at: 07/03/26, 07:53 AM

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

Source: https://huggingface.co/papers/2607.02255

Abstract

A bounded contract approach for long-horizon LLM agents uses typed retrieval to assemble fresh prompts, enabling isolated analysis of memory components and demonstrating improved performance in complex decision-making tasks.

Memory for along-horizon LLM agentis a contract about what each future decision is allowed to see. The simplest contract appends past observations, tool calls, and reflections to every prompt, which makes prior context easy to access but also turns it into a jumbled mixture in which the effect of any single memory component is hard to isolate. We introduce and instrument an alternativebounded contract: every decision is made from a fresh user message assembled bytyped retrieval, with no raw cross-decision transcript appended. The prompt thus stays bounded across runs of any length, and any single layer can be ablated in isolation. We instantiate the contract inSlay the Spire 2, a closed-rule stochastic deck-building game whose runs require hundreds of tactical and strategic decisions. A public online benchmark of frontier LLMs on the same game reports zero wins at the lowest difficulty across five configurations, and the developer-reported human win rate at the same difficulty is 16%; the task is hard but not saturated. Within our harness, a fixed-A0 ablation shows the largest observed difference when triggeredstrategic skillsare enabled: the no-store baseline wins 3/10 games and adding the skill layer 6/10. At this sample size the comparison is directional rather than statistically decisive (Fisher exact p\approx0.37); a cross-backbone probe and public accumulating-context baselines are reported as operational comparisons rather than controlled tests of the contract variable itself. We release a reproducible testbed: 298 completed trajectories with condition tags, frozen memory/skill snapshots, prompt records, and analysis scripts -- an agent design and a validated, reusable methodology for studying howexplicit memorylayers shape long-horizon LLM-agent decisions.

View arXiv pageView PDFProject pageGitHub5Add to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2607.02255 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2607.02255 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2607.02255 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Shared Selective Persistent Memory for Agentic LLM Systems

arXiv cs.AI

This paper introduces shared selective persistent memory for agentic LLM systems, which retains reusable context like task specs and data schemas while discarding session-specific reasoning traces, achieving 96% task completion and significant token cost reductions in enterprise scenarios.

Selective Memory Retention for Long-Horizon LLM Agents

arXiv cs.AI

This paper presents TraceRetain, a lightweight framework for bounded external memory in frozen LLM agents, demonstrating that selective retention differentiates from cache heuristics primarily when memory streams contain noise, offering task-success and efficiency benefits.

Speculate with Memory: Lossless Acceleration for LLM Agents

arXiv cs.LG

This paper introduces memory-augmented speculative execution for LLM agents, using three online memory systems to improve prediction accuracy by 19-39% on action prediction and up to 2.5x on observation prediction, all while being lossless with zero added wall-clock cost.

Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents

Hugging Face Daily Papers

This paper introduces a proactive memory agent that runs alongside an action agent to prevent behavioral state decay in long-horizon tasks, achieving significant improvements on Terminal-Bench2.0 and τ^2-Bench. The authors also train Qwen3.5-27B using SFT and GRPO as an early step toward open-weight memory policies.