Most agent-memory work fixes the wrong half of the problem. It assumes the hard part is storing procedural knowledge and pulling back the closest text, so it pours effort into embeddings and rerankers. CaSKG makes the case that the hard part is the edges — the dependency relations between skills — and that’s the half almost everyone treats as free. Vector retrieval hands you a compact neighborhood but scores each skill as independent text, so it will happily return a completion step without its prerequisite. Graph retrieval can carry that workflow context, but only when the edges are trustworthy, and in practice they’re built from surface similarity and quietly wrong.

CaSKG calibrates edge confidence before any retrieval happens. It builds a high-recall candidate graph, then runs direction-conditioned counterfactual probes — remove, substitute, and reorder skill pairs — to see which relations actually change the task outcome, and smooths that evidence into a weighted graph. The property I care about is operational, not the leaderboard: the graph is constructed offline and used without touching the agent policy or the task interface. That’s the line between a research artifact and something you can drop into a running system.

Calibrated edges preserve prerequisites, state-changing actions, and verification routines — exactly the steps a naive reranker sheds first, and exactly the ones whose absence turns into a silent failure three actions later. The gains back it up (macro-average ScienceWorld climbs from 72.6 to 80.5 over Graph-of-Skills, with fewer environment steps), but the reusable lesson is that edge confidence is a first-class retrieval signal, not metadata you bolt on afterward. The HF paper page has the ablations if you want to see which edge types carry the weight.

If you run a tool or skill library at any real scale, where is your eval budget going — into ranking the nodes, or into trusting the edges that connect them?

tags: [ agentic-ai ] [ rag ] [ knowledge-graphs ] [ research ]