搜索

PythonReactSEOPDF
GitHub Skill

ai-maestro-memory-search

PROACTIVELY search conversation history when receiving user instructions. Find previous discussions, decisions, and context BEFORE starting new work. Your memory is valuable - use it.

aiskillstore☆ 124 仓库 Stars6 仓库 Forks

安装与使用

请根据来源仓库和你的 AI 工具说明安装此技能。先阅读技能文件和依赖要求,再添加到工作流。

查看来源 ↗使用指南

技能说明

AI Maestro Memory Search

CRITICAL: AUTOMATIC BEHAVIOR - READ THIS FIRST

THIS IS NOT OPTIONAL. THIS IS YOUR DEFAULT BEHAVIOR.

When the user gives you ANY instruction, you MUST FIRST search your memory for:

  • Previous discussions - Have we talked about this before?
  • Past decisions - What did we decide last time?
  • Prior context - What does the user expect based on history?
  • Earlier work - Did we already start on this?

DO NOT:

  • Start fresh without checking conversation history
  • Assume this is a new topic without verifying
  • Skip memory search because "I don't remember discussing this"
  • Wait for the user to say "remember when we..."

ALWAYS:

  • Search memory IMMEDIATELY when you receive a task
  • Search for terms and concepts the user mentions
  • Check for previous decisions on similar topics
  • Look for context that informs current work
The Rule: Receive Instruction → Search Memory → Then Proceed
1. User asks you to do something
2. IMMEDIATELY search memory for relevant context
3. NOW you know what was discussed before
4. NOW you can build on previous work, not start over

Available Commands

Command Description
memory-search.sh "<query>" Hybrid search (recommended)
memory-search.sh "<query>" --mode semantic Find conceptually related
memory-search.sh "<query>" --mode term Exact term matching
memory-search.sh "<query>" --role user Only user messages
memory-search.sh "<query>" --role assistant Only your responses

What to Search Based on User Instruction

User Says IMMEDIATELY Search
"Continue working on X" memory-search.sh "X"
"Fix the issue we discussed" memory-search.sh "issue", memory-search.sh "bug"
"Use the approach we agreed on" memory-search.sh "approach", memory-search.sh "decision"
"Like we did before" memory-search.sh "<topic> implementation"
Any specific feature/component memory-search.sh "<feature>"
References to past work memory-search.sh "<reference>" --mode semantic

Quick Examples

# User asks to continue previous work
memory-search.sh "authentication"
memory-search.sh "last session"

# User mentions a component we discussed
memory-search.sh "PaymentService" --mode term

# Find what the user previously asked for
memory-search.sh "user request" --role user

# Find your previous solutions
memory-search.sh "implementation" --role assistant

# Conceptual search for related discussions
memory-search.sh "error handling patterns" --mode semantic

Search Modes

Mode Use When
hybrid (default) General search, best for most cases
semantic Looking for related concepts, different wording
term Looking for exact function/class names
symbol Looking for code symbols mentioned

Why This Matters

Without searching memory first, you will:

  • Repeat explanations the user already heard
  • Contradict previous decisions
  • Miss context that changes the approach
  • Start over instead of continuing

Memory search takes 1 second. Frustrating the user is much worse.

Combining with Doc Search

For complete context, use BOTH:

# User asks about creating a new feature
memory-search.sh "feature"       # What did we discuss?
doc-search.sh "feature"          # What do docs say?

Error Handling

If no results found, that's valuable information too: "No previous discussions found about X - this appears to be a new topic. Let me search the documentation..."

Then search docs as fallback.

Script not found:

  • Check PATH: which memory-search.sh
  • Verify scripts installed: ls -la ~/.local/bin/memory-*.sh
  • Scripts are installed to ~/.local/bin/ which should be in your PATH

Installation

If commands are not found:

./install-memory-tools.sh

This installs scripts to ~/.local/bin/.

使用示例

查看来源仓库中的使用示例与配置说明。

查看来源 ↗

来源历史

收录时间
2026-01-20
索引更新时间
2026-01-28
Branch
main

收录与同步日期属于本站记录。源码修改历史请查看来源仓库。

GitHub ↗

chat Comments (0)

chat_bubble_outline

No comments yet. Be the first to share your thoughts!