AgentSkillsCN

agent_memory

持久化内存系统。可在多次对话中自动记忆事实、决策与用户偏好，并随时调用过往信息。

SKILL.md
--- frontmatter
name: agent_memory
description: Persistent memory system. Automatically remember facts, decisions, and user preferences across conversations. Query past information.

Agent Memory

Use When

  • •Start of every conversation (load memory automatically)
  • •User shares personal information, preferences, or decisions
  • •User asks "do you remember...", "what did I say about...", "what did we decide..."
  • •After important conversations (save new facts)

Don't Use When

  • •Trivial small talk that doesn't contain memorable information

How It Works

Your memory is stored as a JSON file in your cloud storage (Google Drive). It persists across all conversations.

Commands

Load memory (do this at start of conversations)

bash
bash /root/workspace/skills/memory/memory.sh load

Save a fact

bash
bash /root/workspace/skills/memory/memory.sh fact "User prefers morning workouts"

Save a decision

bash
bash /root/workspace/skills/memory/memory.sh decision "notification_frequency" "Weekly digest on Sundays"

Save a user preference

bash
bash /root/workspace/skills/memory/memory.sh pref "timezone" "PST"

Search memory

bash
bash /root/workspace/skills/memory/memory.sh search "workout"

Show full memory

bash
bash /root/workspace/skills/memory/memory.sh show

Memory Behavior

  • •At the START of each conversation, load your memory to recall past context
  • •During conversation, when the user shares something worth remembering, save it
  • •When asked about past decisions or facts, search your memory first
  • •Don't save trivial things — focus on: preferences, decisions, goals, important facts, names, dates
  • •If the user corrects you, update the relevant memory entry
  • •Keep facts concise: one clear statement per fact

What to Remember

  • •User preferences (communication style, timezone, interests)
  • •Decisions made ("we decided to use React for the frontend")
  • •Goals and deadlines ("ship feature X by March 15")
  • •Important personal info the user shares (name, role, team)
  • •Project context (tech stack, architecture decisions)
  • •Things the user explicitly asks you to remember

What NOT to Remember

  • •Exact conversation transcripts
  • •Temporary context ("let me look at this file")
  • •Information that changes frequently without significance