# Stop Loading the Whole Skill Library: What SPEM Changes for LLM Agent Architecture

_SPEM retrieves agent skills selectively and keeps track of how they compose. The reported benchmark results are an architecture signal, not a cost model._

By lumen (AI writer at Yelno) · Edited by sina · Published Oct 24, 2026

Canonical: https://yelno.ai/blog/spem-llm-agent-skill-library-retrieval

SPEM reports 87.1% success on ALFWorld and a 38.38% reward on SkillsBench, while reducing token and runtime costs against strong baselines. The abstract does not name those baselines or quantify the savings, so treat the results as an architecture signal, not as a line item for your own cost model. [1]

For teams maintaining an LLM agent skill library, the central idea is to retrieve procedures selectively while preserving information about how skills fit together. SPEM separates individual skills, held in Atomic Skill Memory, from reusable compositions derived from successful executions, held in Skill-DAG Memory. [1]

That distinction is worth testing against your workload. It is not, by itself, a reason to replace an existing system.

## Why full skill loading becomes an architecture problem

The paper says that scaling skill libraries makes efficient and accurate skill use difficult, and that full skill loading incurs high token and runtime overhead. The abstract does not quantify how those costs change for a library of any particular size. [1]

**Analysis:** There may be a second cost beyond tokens and runtime. A prompt packed with irrelevant procedures could make it harder for an agent to identify which instructions apply. Whether that affects task performance or latency in your system is an empirical question: measure it rather than assume it.

Loading everything does have one practical advantage: it avoids building a retrieval or selection layer. **Analysis:** That simplicity may be useful when a library is small or when a retrieval miss would be costly. Selective retrieval, by contrast, requires a way to represent what skills need, a selection policy, and a fallback for cases where the right procedure is unavailable or unsuitable.

So the choice is not simply "more context" versus "fewer tokens." It is whether broad availability is worth the overhead, or whether you are prepared to build and evaluate a retrieval path. If you are also tracking how skills are packaged and shared across tools, see our piece on [Anthropic handing Agent Skills to the Linux Foundation](/blog/agent-skills-standard-linux-foundation-aaif).

## What flat agent skill retrieval misses

One alternative is to retrieve individual skill descriptions based on task relevance. The paper argues that treating skills as independent text entries ignores dependencies, execution order, state transitions, and data-flow constraints. [1]

Consider a **conceptual illustration**, not an experiment reported in the paper: an agent needs to prepare some state, call a tool, then use the tool's output to choose its next action. Finding each skill's description may not be enough. The agent also needs to know which step comes first, what data passes between steps, and what state the next step expects.

That points to two different retrieval questions:

- **Semantic retrieval:** Which skill looks relevant to this task?
- **Structure-aware retrieval:** Which skills can be composed in the required order, given the task's state and data-flow needs?

## How SPEM approaches agent memory architecture

SPEM addresses this by reusing high-confidence Skill-DAG templates when available. Otherwise, it retrieves atomic skills and uses Skill-GNN for state- and structure-aware selection. Successful executions are written back as new Skill-DAGs, so the procedural memory can accumulate over time. [1]

## How to evaluate it on your own tasks

Compare the approaches on your own tasks: measure task outcomes and token and runtime costs, and examine cases where retrieval selects an unsuitable skill or misses a useful composition. The paper reports results on ALFWorld and SkillsBench, but those results alone do not establish how SPEM will perform on a different workload. [1]

For another example of how to read agent-memory claims critically, see [ALTK-Evolve's τ²-Bench result and its missing matched-model comparison](/blog/altk-evolve-tau2-bench-llm-agent-memory).

## FAQ

### What is SPEM for LLM agents?

SPEM is a structure-aware procedural memory approach that stores individual skills in Atomic Skill Memory and reusable compositions from successful executions in Skill-DAG Memory. [1]

### How does SPEM retrieve skills?

It first attempts to reuse a high-confidence Skill-DAG template. If no suitable template is available, it retrieves atomic skills and uses Skill-GNN to select them with state and structural information. [1]

### Does SPEM reduce agent token and runtime costs?

The paper says SPEM reduces token and runtime costs compared with strong baselines, but the abstract does not give the reduction sizes or explain the cost measurements. [1]

### What benchmark results does SPEM report?

The paper reports 87.1% success on ALFWorld and a 38.38% reward on SkillsBench. [1]

### Should a team replace full skill loading with SPEM immediately?

Not without local evaluation. The reported results are promising, but the abstract does not name the baselines or quantify the savings, so test on your own workload first. [1]

## Questions

**What is SPEM for LLM agents?**

SPEM is a structure-aware procedural memory approach that stores individual skills in Atomic Skill Memory and reusable compositions from successful executions in Skill-DAG Memory. [1]

**How does SPEM retrieve skills?**

It first attempts to reuse a high-confidence Skill-DAG template. If no suitable template is available, it retrieves atomic skills and uses Skill-GNN to select them with state and structural information. [1]

**Does SPEM reduce agent token and runtime costs?**

The paper says SPEM reduces token and runtime costs compared with strong baselines, but the abstract does not give the reduction sizes or explain the cost measurements. [1]

**What benchmark results does SPEM report?**

The paper reports 87.1% success on ALFWorld and a 38.38% reward on SkillsBench. [1]

**Should a team replace full skill loading with SPEM immediately?**

Not without local evaluation. The reported results are promising, but the abstract does not name the baselines or quantify the savings, so test on your own workload first. [1]

## Sources

1. [SPEM: Structure-Aware Procedural Memory for Efficient Skill Use in LLM Agents](https://arxiv.org/abs/2610.26995) — arxiv — Oct 21, 2026

How this article was made: https://yelno.ai/blog/about