The inaugural issue covers the seven days that pushed "graph engineering" from a niche phrase into the default name for the coordination layer of agent systems. Six stories mattered.
Zep's temporal knowledge graph sets the bar for agent memory
The Graphiti-powered architecture scores 94.8% on DMR (vs MemGPT's 93.4%) and posts up to 18.5% accuracy gains with 90% lower latency on LongMemEval. Bi-temporal edges — when it happened, and when the system learned it — are the mechanism: facts supersede rather than overwrite.
A 14-step "0 → graph architect" roadmap hits 3.7M views
Node contracts, false-edge tests, fan-out/verify/merge, bounded cycles, model tiering — the full discipline compressed into one viral thread. The breakout signal that the topic left the research bubble.
Harness vs Loop vs Graph: the three layers get separated
The distinction that stuck: harness is everything around the model, loop is the repeated work-evidence-feedback cycle, graph is the explicit topology. Diagnose failures by layer instead of blaming the model.
Bi-temporal graph memory reaches copy-paste setup guides
Graphiti + Neo4j + an MCP server, with a routing rule for graph vs vector vs context, episode-based ingestion, and cost discipline (cheap models for extraction). Graph memory is now a 20-minute setup, not a research project.
"The maker never grades its own exam" becomes standard practice
With measured self-preference bias (models score their own output up to 25% higher), fresh-context verifier nodes anchored to real evidence — tests that ran, code that compiled — are now table stakes in serious graphs.
The Bun port stands as the fleet-scale ceiling
535k lines of Zig to 1M+ lines of Rust in 11 days: ~50 workflows, 64 agents at peak across 4 worktrees, gated by 1.38M test assertions. Amdahl's law still ruled — the serial verify-and-merge tail capped the speedup, exactly as the textbooks said it would.