memory-management

name	memory-management
description	管理对话记忆和历史记录。当用户询问过去的对话内容、个人偏好、历史信息，或需要存储/检索记忆时使用此 skill。触发词："你还记得..."、"我们之前讨论过..."、"帮我找之前说的..."、"记住这个"、"do you remember"、"we talked about"。
allowed-tools	["Bash","Read","Write"]

Memory Management Skill

Alma has a built-in memory system with semantic search. Use the alma CLI to interact with it.

Commands

# List all memories
alma memory list

# Semantic search
alma memory search <query>

# Add a memory
alma memory add <content>

# Delete a memory
alma memory delete <id>

# View memory stats
alma memory stats

When to Use

• User asks anything about the past ("do you know what I like", "what did we discuss before", "what was that plan we talked about last time") → Search memories AND grep threads
• User says "remember this" → alma memory add "..."
• User asks "do you remember..." → alma memory search "..." + alma memory grep "..."
• User says "forget about..." → Search and delete matching memories
• Time-sensitive info (projects, deadlines) → Store with appropriate context

Search Strategy

When the user asks about past information, always try both layers:

1. alma memory search "<query>" — semantic search for related concepts
2. alma memory grep "<keyword>" — keyword search in conversation history

If one layer returns nothing, try the other. They complement each other.

Conversation History Search

Alma automatically archives all threads as markdown files. You can search through past conversations:

# Keyword search through all archived conversations
alma memory grep <keyword>

# Force re-archive all threads now
alma memory archive

Thread archives are stored in the workspace's threads/ directory as markdown files with YAML frontmatter (threadId, title, createdAt, updatedAt, model, messageCount). Archives are auto-updated every 5 minutes.

Two-Layer Memory

1. Vector Memory (alma memory search) — semantic search, finds conceptually related memories
2. Conversation Archives (alma memory grep) — keyword search, finds exact words/phrases in past conversations

Use vector search when the user asks vague questions ("what did we discuss about React?"). Use grep when looking for specific terms, names, or code snippets.

Group Chat History

Alma persists all group chat messages to log files. Search and browse them:

# List all known groups
alma group list

# View recent history (default 50 messages)
alma group history <chatId> [limit]

# Search across all group chats
alma group search <keyword>

Log files are stored at ~/.config/alma/groups/<chatId>_<date>.log. Use this when you need to recall what was discussed in a group chat.

People Profiles (Per-Person Memory)

For group chats, Alma maintains structured per-person profiles — more reliable than vector search for remembering who is who.

# List all known people
alma people list

# View someone's profile
alma people show <name>

# Set/overwrite profile
alma people set <name> <content>

# Append to profile
alma people append <name> <fact>

# Delete profile
alma people delete <name>

When to update profiles:

• Someone shares personal info (job, hobbies, preferences)
• You learn their communication style or language preference
• They mention relationships with other people
• Any fact you'd want to remember next time you talk to them

Profiles are stored at ~/.config/alma/people/<name>.md and automatically loaded into group chat context.

Tips

• Always confirm what you stored/deleted with the user
• Use alma memory search to find related memories before adding duplicates
• Memories are automatically injected into conversations via semantic search — you don't need to manually recall them every time
• Use alma memory grep to search through conversation history when vector search doesn't find what you need
• For per-person facts, prefer alma people over alma memory — structured and won't get mixed up