How to have Claude Code run 28× longer before auto-compaction (github.com)

A new technique called Context Ledger has been introduced to enhance the performance of long-horizon coding agents by optimizing their memory usage during code development. Traditionally, as agents accumulate context information while building features, they risk exceeding their token limit, leading to either failures or loss of essential details. Context Ledger addresses this by compressing the build context at natural commit boundaries using structured ledger entries that retain key information while linking back to the original git commit. This allows agents to effectively "evacuate" less critical data while still being able to recover necessary details from git when needed. The significance of Context Ledger lies in its ability to sustain the agent's operational efficiency significantly longer—up to 28 times longer—before hitting the memory limit. In practical tests, it maintained a high retention rate (93%) of relevant information about earlier features while reducing working context size substantially compared to traditional methods. This advancement not only preserves crucial semantic details but also facilitates a lossless memory management process, allowing for more efficient coding workflows. The implementation offers a comprehensive CLI for seamless integration with existing projects, making it a valuable tool for developers in the AI/ML community.