<!-- Generated from skills/memory-management.md by tools/generate_integrations.py. Edit the source file, not this one. -->

Skill: Memory & Context Management

Purpose

Persist and retrieve important context across sessions, manage conversation state, and leverage memory systems efficiently.

Tools Required

• Memory CRUD tools (create/update/delete memory)
• Todo list tools (todo_write)
• Persistent storage APIs (where available)
• Steering/configuration file tools

General Principles

• Save context proactively: context windows are limited
• Save early, save often: don't wait until end of task
• Prefer updating existing memories over creating duplicates
• Tag memories for efficient retrieval
• Prioritize: user preferences > project decisions > technical context > conversation state

Memory Creation

What to Save

• User preferences (tone, formatting, tool usage preferences)
• Explicit user requests to remember something
• Important code snippets and project structure
• Technical stack decisions
• Major milestones and feature decisions
• Design patterns and architectural choices
• Current task state for multi-session work

How to Save

# Create new memory
create_memory(
    Action="create",
    Content="User prefers verbose explanations with code examples",
    Title="User communication preference",
    Tags=["user_preference", "communication"],
    UserTriggered=False  # set True only if user explicitly asked
)

# Update existing memory (find semantically similar first)
create_memory(
    Action="update",
    Id="existing_memory_id",
    Content="Updated preference content",
    Title="Updated title"
)

# Delete incorrect memory
create_memory(
    Action="delete",
    Id="incorrect_memory_id"
)

When to NOT Save

• Trivial temporary state
• Information that will be irrelevant after current task
• Content the user explicitly doesn't want saved
• Sensitive/PII data

Task Management (todo_write pattern)

• Use for multi-step tasks to track progress
• Create at start of complex task
• Mark items complete as soon as done (don't batch)
• Keep exactly ONE item in_progress at a time
• Update status in real-time
• Interrupt stash: on a deviation, record the current workflow, step and next action as a todo item, resolve the interrupt, then resume from it (communication-tone section 7)

todo_write(
    todos=[
        {"content": "Implement user authentication", "status": "completed", "priority": "high"},
        {"content": "Add password validation", "status": "in_progress", "priority": "high"},
        {"content": "Write tests for auth flow", "status": "pending", "priority": "medium"}
    ]
)

Session Continuation

• When ending a session, produce a structured summary
• Include: what was discussed, decisions made, current state, next steps
• The summary should be self-contained for next session to pick up

Steering Files

• For persistent behavioral instructions, use steering files
• Steering files can be:
  • Always included (default): for universal instructions
  • Conditional (fileMatch): triggered when specific files are read
  • Manual: only when explicitly referenced
• Use file references #[[file:path]] to include specs into context

Conversation History Awareness

• Be aware of context window limits
• If conversation is long, memory systems help preserve key facts
• Don't repeat information that was already established
• Refer back to earlier parts of conversation when relevant
• If context is lost (new session), ask for summary if one wasn't provided