7.7
backpass Review 2026 — Gradient Descent for Your Agent Memory (AGENTS.md as Weights)
backpass treats your AGENTS.md as a set of weights: it reads local session transcripts from seven agent harnesses, computes which instructions actually helped or were violated, and proposes evidence-backed, budgeted edits you approve one by one. We review the collect-loss-gradient-apply pipeline, the 5,000-token budget gate, skills-as-overflow, and the no-DEFER apply UI.