Ex Memory Distiller

Problem

When users want to turn chat logs, photos, and social posts into a local ex-partner dialogue skill, a plain agent often produces only a shallow summary. It is hard to preserve speech habits, relationship context, and later corrections. ex-friend treats the output as maintainable local files instead of one-off generation.

How it works

• Intake: collect only an alias, basic relationship info, and personality profile.
• Source import: parse WeChat/QQ exports, social screenshots, photo EXIF data, PDF/TXT files, or pasted notes, using Read, Bash, and local scripts to extract catchphrases, reply cadence, topic distribution, timeline, and relationship events.
• Two-layer build: write memory.md for shared experiences, habits, conflict patterns, and inside jokes; write persona.md for speech style, emotional expression, and behavioral boundaries; then generate a runnable SKILL.md.
• Iteration: support appending memory, correcting “they would not say that,” rolling back versions, and deletion, so the simulated persona stays tied to source evidence rather than idealized invention.

Boundaries

This skill is intended for personal reflection and emotional healing, with data kept local. It should not be used for harassment, stalking, replacing real communication, or encouraging unhealthy attachment. The generated persona must follow Layer 0 rules: it should not suddenly confess, apologize, or become unconditionally accepting without supporting material.