late-night-regrets

late-night-regrets is an extension for PiClaw that provides nightly analysis of chat interactions using a Bayesian classifier. Designed to help users reflect on the quality of their conversational exchanges, it processes chat message histories stored in SQLite databases, identifying behavioral patterns and generating self-improvement reflections. " by training on past conversations and flagging attention-worthy messages such as corrections, misinterpretations, over-engineering, under-delivery, and context failures.

The classifier operates using a multinomial Naive Bayes approach, relying on tokenized text and structural features derived from the flow of conversation, such as previous sender, message length, and repetition. It is weakly supervised, meaning it detects patterns directly from the text without requiring manual labeling. Classification is performed without consuming model tokens, reserving model-based reasoning only for the flagged subset of messages during the reflection pass. Output from each nightly run includes various artifact files, such as interaction-quality weights, predictions, attention-worthy messages, and a report in markdown format. Reflections and feedback are written to specific files within the user's notes and memory directories, and these integrate with Dream's standard memory hierarchy.

late-night-regrets can be installed through PiClaw's Add-Ons settings, with configurable options such as enabling/disabling the extension, setting the nightly cron schedule, adjusting the confidence threshold for flagging messages, and specifying output paths. Users can also manually trigger a reflection pass with a command. The classifier labels user messages following agent turns into categories like successful execution, course correction, misinterpretation, over-engineering, under-delivery, context failure, good proactive, and neutral.

tgz package. It is intended for users of PiClaw interested in automated, nightly interaction quality assessment and self-improvement insights based on their chat history.