SPEM Hits 87.1% ALFWorld Success via Skill-DAG Memory
SPEM combines atomic skill storage with DAG-based procedural memory and a structure-aware GNN selector. On ALFWorld and SkillsBench it delivers the strongest effectiveness-efficiency frontier reported. The framework supplies an explicit path for agents to accumulate reusable execution graphs rather than isolated skill text.
The October 2026 arXiv paper 2610.26995 presents SPEM as a dual-memory system. Atomic Skill Memory stores individual skills. Skill-DAG Memory records verified execution graphs from prior tasks. For novel queries the system first checks for matching high-confidence DAG templates; absent a match it falls back to atomic retrieval routed through a Skill-GNN that encodes state transitions and data-flow constraints.
ALFWorld and SkillsBench results show the efficiency gain. Full skill loading and semantic-only retrieval both incur higher token counts and lower success. SPEM reuses DAG templates on repeat patterns, reducing context size and avoiding redundant planning steps. The write-back loop incrementally expands the DAG library from successful trajectories.
Operational impact centers on scaling. Flat retrieval ignores ordering and preconditions; SPEM encodes them explicitly. This pattern matches earlier work on hierarchical task networks in robotics and on compositional generalization benchmarks in language models. The approach therefore supplies a concrete mechanism for accumulating procedural knowledge without proportional growth in inference cost.
Next steps include integration into open agent runtimes and measurement on longer-horizon environments where DAG reuse frequency rises. No production deployment data yet exists.
Agent frameworks: SPEM-style DAG reuse will exceed 70% of planning steps in production agent traces within 18 months of open-source release.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2610.26995)
- [2]Supporting Source(https://arxiv.org/abs/2010.03768)