Trajectory Memory Ledger

We present the Trajectory Memory Ledger, implemented in KARL, a schema-normalized experience replay system for improving AI coding agent performance through closed-loop feedback. The ledger records complete tool-use sequences during real coding sessions, normalizes them into an append-only schema, scores them using a six-signal composite reward function (outcome, process, efficiency, verification, consistency, and wasted motion), and uses the highest-scoring trajectories to generate advantage-weighted supervised fine-tuning data. Unlike approaches that rely on static benchmarks or human preference labels, the Trajectory Memory Ledger derives training signal from observable agent behavior and implicit user feedback. The current normalized deployment corpus contains 7,468 scored trajectories, 67,409 observed tool events, and 73,470 recovered tool steps across 50+ active projects. From this store, KARL exports 3,678 ChatML training examples (3,310 train / 368 validation). We describe the system architecture, schema normalization, reward design, OAPL-Lite export, and entity bridge for performance-based skill decay.