Agent Skills: Searching Messages

Search past messages to recall context. Use when you need to remember previous discussions, find specific topics mentioned before, pull up context from earlier in the conversation history, or find which agent discussed a topic.

UncategorizedID: letta-ai/letta-code/searching-messages

Repository

letta-aiLicense: Apache-2.0
992125

Install this agent skill to your local

pnpm dlx add-skill https://github.com/letta-ai/letta-code/tree/HEAD/src/skills/builtin/searching-messages

Skill Files

Browse the full folder contents for searching-messages.

Download Skill

Loading file tree…

src/skills/builtin/searching-messages/SKILL.md

Skill Metadata

Name
searching-messages
Description
Search past messages to recall context. Use when you need to remember previous discussions, find specific topics mentioned before, pull up context from earlier in the conversation history, or find which agent discussed a topic.

Searching Messages

This skill helps you search through past conversations to recall context that may have fallen out of your context window.

When to Use This Skill

  • User asks "do you remember when we discussed X?"
  • You need context from an earlier conversation
  • User references something from the past that you don't have in context
  • You want to verify what was said before about a topic
  • You need to find which agent discussed a specific topic (use with finding-agents skill)

CLI Usage

letta messages search --query <text> [options]

Options

| Option | Description | |--------|-------------| | --query <text> | Search query (required) | | --mode <mode> | Search mode: vector, fts, hybrid (default: hybrid) | | --start-date <date> | Filter messages after this date (ISO format) | | --end-date <date> | Filter messages before this date (ISO format) | | --limit <n> | Max results (default: 10) | | --all-agents | Search all agents, not just current agent | | --agent <id> | Explicit agent ID (overrides LETTA_AGENT_ID) | | --agent-id <id> | Alias for --agent |

Search Modes

  • hybrid (default): Combines vector similarity + full-text search with RRF scoring
  • vector: Semantic similarity search (good for conceptual matches)
  • fts: Full-text search (good for exact phrases)

Companion Command: messages list

Use this to expand around a found needle by message ID cursor:

letta messages list [options]

| Option | Description | |--------|-------------| | --after <message-id> | Get messages after this ID (cursor) | | --before <message-id> | Get messages before this ID (cursor) | | --order <asc\|desc> | Sort order (default: desc = newest first) | | --limit <n> | Max results (default: 20) | | --agent <id> | Explicit agent ID (overrides LETTA_AGENT_ID) | | --agent-id <id> | Alias for --agent |

Search Strategies

Strategy 1: Needle + Expand (Recommended)

Use when you need full conversation context around a specific topic:

  1. Find the needle - Search with keywords to discover relevant messages:

    letta messages search --query "flicker inline approval" --limit 5
    
  2. Note the message_id - Find the most relevant result and copy its message_id

  3. Expand before - Get messages leading up to the needle:

    letta messages list --before "message-xyz" --limit 10
    
  4. Expand after - Get messages following the needle (use --order asc for chronological):

    letta messages list --after "message-xyz" --order asc --limit 10
    

Strategy 2: Date-Bounded Search

Use when you know approximately when something was discussed:

letta messages search --query "topic" --start-date "2025-12-31T00:00:00Z" --end-date "2025-12-31T23:59:59Z" --limit 15

Results are sorted by relevance within the date window.

Strategy 3: Broad Discovery

Use when you're not sure what you're looking for:

letta messages search --query "vague topic" --mode vector --limit 10

Vector mode finds semantically similar messages even without exact keyword matches.

Strategy 4: Find Which Agent Discussed Something

Use with --all-agents to search across all agents and identify which one discussed a topic:

letta messages search --query "authentication refactor" --all-agents --limit 10

Results include agent_id for each message. Use this to:

  1. Find the agent that worked on a specific feature
  2. Identify the right agent to ask follow-up questions
  3. Cross-reference with the finding-agents skill to get agent details

Tip: Load both searching-messages and finding-agents skills together when you need to find and identify agents by topic.

Search Output

Returns search results with:

  • message_id - Use this for cursor-based expansion
  • message_type - user_message, assistant_message, reasoning_message
  • content or reasoning - The actual message text
  • created_at - When the message was sent (ISO format)
  • agent_id - Which agent the message belongs to