When Does Sub-Graph Paging Help? An Empirical Comparison of Full-Context, Top-K, Personalized PageRank, and Summary-Indexed Graph Memory for Multi-Agent Systems

I present an empirical study and reproducible in-process benchmark suite evaluating four competing retrieval paradigms for multi-agent software engineering memory under identical compact tuple serialization (25 tokens/node) across five scale tiers (20 to 800 nodes, 300 multi-hop causal path queries): Full-Context Serialization, Flat Top-K Lexical/Vector Retrieval, Budget-Matched HippoRAG (Global Personalized PageRank), and Summary-Indexed Sub-Graph Paging (both naive EngramGraph v1 and anchor-augmented EngramGraph v2). Our evaluation yields four actionable findings:1. Small Graphs (<= 144 nodes / <= 3.6k compact tokens): Full-Context serialization achieves 100% path and node coverage with zero routing overhead; sub-graph paging is unnecessary at this scale.2. Failure Mode of Naive Summary Paging (EngramGraph v1): Routing solely on 2-sentence sub-topic summaries degrades from 95.0% at 20 nodes to 58.3% at 384 nodes and 30.0% at 800 nodes due to summary compression loss and unmounted cross-topic bridges.3. Structural & Anchor Mitigation (EngramGraph v2): Augmenting summary routing with an Entity-Anchor Footer (+12 tok/sub-topic), 1-Hop Bidirectional Bridge Closure, and Anchor-Seeded Subgraph Projection recovers complete-chain recall to 86.7% at 384 nodes (95.6% node coverage) and 88.3% at 800 nodes (96.1% node coverage) while reducing active prompt tokens by 77.2% to 87.7% compared to compact Full Context, and outperforming equal-budget Flat Top-K by +20.0 to +26.6 percentage points (McNemar p < 0.001).4. Read-Only Global PPR vs. Read/Write Sub-Topic Isolation: Budget-Matched HippoRAG (Global Personalized PageRank) achieves the highest read-only recall on static graphs (98.3% at 384 nodes; 93.3% at 800 nodes), whereas EngramGraph v2 provides partitioned sub-topic write leases and Optimistic Concurrency Control (OCC) for concurrent multi-agent mutation workloads with zero silent lost updates.