trajectory-garden

Remove redundant agent trajectories by analyzing JSONL logs to identify duplicate queries and high-cost interaction turns, generating a sourced work-list of prune candidates for human approval to optimize memory stores.

Agent memory retrieval degrades due to the accumulation of redundant and low-quality agent trajectories that bloat memory stores and inflate retrieval costs, an issue addressed by identifying and pruning redundant interaction turns.

Simhash similarity detects duplicate queries in agent trajectory logs by computing lexical similarity scores across interaction turns, allowing the system to automatically rank agent turns based on redundancy and flag them for removal.

You can use custom semantic embedding models for trajectory pruning by utilizing the extensible framework's pluggable seam, which allows developers to swap the reference simhash scorer with custom semantic embedding models.

Pruning agent memory does not delete data automatically; the system generates a concrete, sourced work-list of prune candidates strictly for human approval, ensuring no data is deleted without oversight.

The best way to identify high-cost interaction turns in agent observability logs is to utilize custom scoring hooks that analyze JSONL trajectory corpora to detect performance outliers and rank turns based on cost and denial rates.