AI Agent Hub
Back to skills
Memory Evolution icon

Memory Evolution

AI Agent Updated 2026.08.30

Paste the following prompt into your AI chat to install this skill:

Please install @user_0327da8d/memory-evolution according to https://skillhub.cn/install/skillhub.md.

About this skill

Problem addressed

Agents often repeat the same mistakes, forget user corrections, or keep stale rules in context. This skill turns task experience into local, searchable, decaying memory with explicit promotion and archival rules.

How it works and limits

  • Triggers: automatic reflection after significant work, manual runs via reflect, 回顾, or 做梦, and scheduled dream cycles for decay scoring.
  • Memory tiers: observations go to reflections/log.md, user corrections go to hot/corrections.md, repeated critical patterns may be promoted to hot/memory.md, while low-scoring entries move to warm/ or cold/archive/.
  • Decay model: scores in engine/decay-scores.json control availability; 1.0-0.5 is active, 0.5-0.2 loads on demand, and 0.2-0.05 is searchable only when explicitly requested. Invalid scores reset with a logged warning.
  • Safety: writes to hot/memory.md require confirmation, archival never deletes, conflicts resolve by specificity, recency, and user corrections, and the skill does not access calendar, email, or contacts, make network requests, or infer preferences from silence.

Use Cases

  • After a multi-file refactor, have the agent reflect on failure patterns and log user corrections to memory.
  • While maintaining a recurring agent workflow, use reflect to extract 2-4 reusable rules from the last 7 days.
  • During a heartbeat job, run decay scoring, demote low-score memories, and archive cold entries.
  • When the agent repeatedly violates a rule, confirm the user correction replaces it and leave a trace in logs.

Best For

  • Engineers maintaining agent workflows who want repeated corrections persisted as searchable memory.
  • Builders of multi-step task automation who want agents to log non-obvious lessons after completion.
  • Teams running heartbeat-scheduled agents who want stale memories demoted by score, not retained indefinitely.
  • Product developers needing local memory isolation without email, calendar, contacts, or network access.