AuthorsRuiyao Xu, Tiankai Yang, Wei-Chieh Huang
TL;DR
HYPERSKILL uses hypergraph-structured skill memory with dual-path retrieval and structure-informed maintenance to reach 62.00% SR on xBench with GPT-4o (+3.00 over PlugMem).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing experiential memories store trajectories, workflows, or skills as isolated entries, discarding compositional relationships among subtasks and reusable skills.
Most systems rely on flat vector similarity and unconditional accumulation, so agents retrieve noisy, unstructured memories that limit self-evolving performance on complex multi-step tasks.
HOW IT WORKS
HYPERSKILL’s core mechanism combines Nodes, Hyperedges, Dual-path retrieval, Co-occurrence ranking, and HYPERSKILL Memory Evolving into a single hypergraph-structured memory.
You can think of HYPERSKILL like a card catalog where each card links a whole episode’s subtasks and skills, instead of storing isolated pages in a filing cabinet.
This hypergraph design lets HYPERSKILL reuse procedural structure and quality-weighted skills far beyond what a plain context window or flat vector store can capture.
DIAGRAM
This diagram shows how HYPERSKILL processes a new task using subtask and trajectory paths, then ranks skills by co-occurrence for execution.
DIAGRAM
This diagram shows how HYPERSKILL is evaluated with two backbones and multiple memory baselines on three benchmarks.
PROCESS
Task decomposition and subtask path
HYPERSKILL uses Task decomposition and subtask path to turn the task description dq into a fine grained plan P0 aligned with subtask nodes Vu.
From Subtasks to Trajectories Hyperedge Retrieval
Using From Subtasks to Trajectories Hyperedge Retrieval, HYPERSKILL matches P0 and dq against hyperedge embeddings he to form the fused set E star.
From Trajectories to Skills Skill Retrieval
In From Trajectories to Skills Skill Retrieval, HYPERSKILL aggregates skill nodes from E star and applies Co-occurrence ranking via κ v to select Sq.
HYPERSKILL Memory Evolving
During HYPERSKILL Memory Evolving, HYPERSKILL applies quality driven pruning and structure informed merging using the co occurrence matrix W and propagated embeddings Z tilde.
KEY CONTRIBUTIONS
Hypergraph structured experiential memory
HYPERSKILL introduces Nodes and Hyperedges where each trajectory hyperedge groups subtasks Vu and skills Vs with a distilled lesson and utility score γ, preserving n ary associations.
Dual path retrieval and co occurrence ranking
HYPERSKILL combines Dual-path retrieval over subtasks and trajectories with Co-occurrence ranking, selecting skills by κ v across retrieved hyperedges instead of pure embedding similarity.
Structure informed maintenance for self evolving agents
HYPERSKILL’s HYPERSKILL Memory Evolving performs quality driven pruning with τprune=0.2 and structure informed merging using quality weighted co occurrence matrix W and propagation depth L.
RESULTS
SR xBench GPT 4o
62.00%
+3.00 over PlugMem
SR GAIA GPT 4o
44.24%
+2.30 over PlugMem
SR WebWalkerQA Qwen3 30B
50.59%
+4.71 over MemP
Avg Steps xBench GPT 4o
5.27 steps
matches ExpeL with higher SR
On xBench, GAIA, and WebWalkerQA, which test tool augmented planning and web navigation, HYPERSKILL consistently raises success rate while keeping steps and tool calls competitive with baselines. These numbers show that hypergraph structured skill memory yields higher task success without inflating interaction cost.
BENCHMARK
On xBench, GAIA, and WebWalkerQA, which test tool augmented planning and web navigation, HYPERSKILL consistently raises success rate while keeping steps and tool calls competitive with baselines. These numbers show that hypergraph structured skill memory yields higher task success without inflating interaction cost.
BENCHMARK
Success Rate SR on xBench for HYPERSKILL and representative memory baselines using GPT 4o.
BENCHMARK
Success Rate SR on GAIA for HYPERSKILL and representative memory baselines using GPT 4o.
KEY INSIGHT
HYPERSKILL achieves its best performance with a very small retrieval budget, with k=2 giving 62.0% SR on xBench and 51.2% on WebWalkerQA.
This is surprising because many assume more retrieved memories always help, but HYPERSKILL shows that larger k mainly injects noise and degrades success rate.
WHY IT MATTERS
HYPERSKILL unlocks self evolving agents that reuse compositional workflows and skills through hyperedges, instead of hoarding flat, noisy trajectory logs.
Builders can now deploy long lived agents whose memory stays compact, high quality, and structurally aware, enabling continual improvement across diverse web and tool use tasks.
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…