A new arXiv paper maps agent memory, skills, and rules onto a single Experience Compression Spectrum ranging from 5-20× episodic memory to 1,000×+ declarative rules, revealing near-zero cross-community citations and the absence of adaptive mechanisms that shift compression levels on demand (Zhang, 2026; Packer et al., 2023).

The primary source documents that 20+ surveyed systems lock to predetermined compression ratios, independently re-solving retrieval latency and context-window problems without exchanging solutions; evaluation protocols remain coupled to specific compression tiers while transferability rises with compression at the expense of episodic specificity (Zhang, 2026). MemGPT demonstrates contextual memory paging at lower compression (Packer et al., 2023), whereas Voyager's skill library operates at mid-spectrum procedural compression, confirming the paper's shared-subproblem thesis yet exposing its under-emphasis on explicit lifecycle management of compressed knowledge across sessions (Wang et al., 2023).

This spectrum directly addresses core fragmentation by treating memory, skills, and rules as points on an experience-compression axis, aligning with the larger trajectory toward coherent LLM reasoning architectures that compress interaction traces into reusable, adaptive structures; the identified "missing diagonal" of cross-level transitions supplies a concrete design principle for next-generation agents capable of long-horizon, multi-session coherence beyond what isolated memory or skill papers currently target (Zhang, 2026; Packer et al., 2023; Wang et al., 2023).