AuthorsXiangchen Cheng, Yunwei Jiang, Jianwen Sun et al.
TL;DR
AgenticSTS uses a bounded, typed L1–L5 memory contract to keep prompts fixed-size and directionally lifts A0 wins from 3/10 to 6/10 with triggered skills.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Public Slay the Spire 2 benchmarks report zero A0 wins across five frontier LLM configurations, while humans win 16% at the same difficulty.
In this regime, transcript-appending agents face unbounded context growth, making individual memory components un-attributable and long-horizon decisions hard to debug or improve.
HOW IT WORKS
AgenticSTS builds every decision from five typed layers: L1 protocol instructions, L2 state-typed prompts, L3 game knowledge, L4 episodic memory, and L5 skill library.
You can think of AgenticSTS like a computer separating OS rules, schemas, a read-only game database, a case log, and a playbook instead of dumping a giant terminal log into RAM.
This typed contract lets AgenticSTS cap prompt size, toggle each layer independently, and attribute behavior changes to specific memory slots in ways a plain context window cannot.
DIAGRAM
This diagram shows how AgenticSTS composes a fresh user message from L1–L5 for each decision instead of appending a transcript.
DIAGRAM
This diagram shows how AgenticSTS evaluates five memory configurations at A0, isolating prompt strictness, L5 skills, and L4 episodes.
PROCESS
Per-decision compositional context
AgenticSTS receives the current game state and uses L1 protocol instructions plus L2 state-typed prompts to define the decision schema and legal actions.
Five typed knowledge layers
AgenticSTS queries L3 game knowledge for rules, L4 episodic memory for past summaries, and L5 skill library for triggered guides relevant to the state.
Routing and combat truncation
AgenticSTS routes the decision to fast, strategic, analysis, or evolution tiers and truncates combat context to a small conversation object instead of a full transcript.
Skill discovery and Mode B
After runs, AgenticSTS can perform mistake-driven skill discovery and Mode B stub-template-filled authoring to update the L5 skill library under a strict write gate.
KEY CONTRIBUTIONS
Per-decision composition interface
AgenticSTS defines a per-decision composition interface that assembles prompts from L1 protocol instructions through L5 skill library instead of a raw transcript, making each layer individually ablatable.
Evidence for triggered L5 skills
AgenticSTS shows that enabling L5 skill library directionally lifts A0 wins from 3/10 to 6/10 in the fixed-A0 matrix, with Wilson 95% CIs [10.8%, 60.3%] and [31.3%, 83.2%].
Reusable 298-trajectory archive
AgenticSTS releases 298 completed trajectories with condition tags, frozen L4 episodic memory and L5 skill library snapshots, prompt records, and Wilson/bootstrap scripts for reproducible long-horizon memory studies.
RESULTS
Win
6/10
+3/10 over baseline-strict
Win
3/10
baseline-strict no scaffold
Score
82.1
full-frozen skills+episodes at A0
Score
70.4
baseline-strict at A0
On fixed A0 Slay the Spire 2 runs, AgenticSTS evaluates five memory configurations and finds the largest observed difference when L5 skill library is enabled, directionally improving wins from 3/10 to 6/10 under the bounded contract.
BENCHMARK
On fixed A0 Slay the Spire 2 runs, AgenticSTS evaluates five memory configurations and finds the largest observed difference when L5 skill library is enabled, directionally improving wins from 3/10 to 6/10 under the bounded contract.
BENCHMARK
Win counts across AgenticSTS memory configurations at A0.
KEY INSIGHT
AgenticSTS shows that template-filled Mode B skills reach the same 6/10 A0 win rate as hand-authored Mode A skills, despite simpler prose.
This challenges the assumption that sophisticated human-written strategy text is necessary; the key seems to be having a structured L5 skill library, not its exact wording.
WHY IT MATTERS
AgenticSTS makes memory a first-class, ablatable interface, letting researchers toggle L3 game knowledge, L4 episodic memory, and L5 skill library independently on a hard long-horizon game.
Builders can now run controlled experiments on memory contracts, compare bounded versus accumulating context in one harness, and ship agents whose long-horizon behavior is auditable and reproducible.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…