KARL - Trajectory Memory Ledger

KARL is the reference implementation of the Trajectory Memory Ledger: a schema-normalized experience replay layer for AI coding agents. It records what an agent does during real work sessions, normalizes those recordings into a schema-v2 trajectory store, scores them with a six-signal reward engine, and uses the highest-scoring trajectories to improve future performance through LoRA fine-tuning and learned skill routing. Based on the Trajectory Memory Ledger paper ([arXiv 2603.05218](https://arxiv.org/abs/2603.05218)). The trajectory store grows over time. The current normalized store contains 7,468 scored trajectories, 67,409 observed tool events, and 73,470 recovered tool steps. The best trajectories are exported as advantage-weighted SFT data and used to train a LoRA adapter via MLX; the current export contains 3,678 ChatML examples split into 3,310 train and 368 validation rows. | Signal | Weight | Measures | |--------|--------|----------| | **Outcome** | 25% | Was the user satisfied? No correction, no redo, build passed, session continued | | **Process** | 22% | Did tools work? Success rate, bash exit codes, error density | | **Efficiency** | 13% | Was it efficient? Tool diversity, duration efficiency, file touch rate | | **Verification** | 13% | Did the agent check its own work? Tests, builds, read-after-write | | **Consistency** | 13% | Was the tool order coherent? Low thrash, enough context before mutation | | **Wasted motion** | 14% | Did the trajectory avoid repeated loops and unnecessary retries? | Composite reward is `[0, 1]`. Advantage = reward - domain baseline, used for OAPL-Lite oversampling. On the normalized corpus, mean reward is 0.6632 and the reward range is 0.4666-0.8165.