Mem Search

What it does

The Mem Search skill enables an AI agent to retrieve relevant information from past conversations and external knowledge sources, specifically focusing on previous work across all sessions. It follows a three-step workflow: search for indexed IDs, filter results using timelines, and then fetch the full details of only the most pertinent items. This allows the agent to provide contextually aware responses based on prior interactions and project history.

When to use it

This skill is best used in scenarios where you need information from past sessions, such as:

• When a user asks about previously addressed issues ("Did we already fix this?").
• To recall how similar problems were solved in the past ("How did we solve X last time?").
• To understand what transpired during previous interactions ("What happened last week?").
• For quickly referencing project history and decisions.

Key capabilities

• Search: Retrieves a list of indexed IDs based on a search query, limit, and project filter.
• Timeline: Provides context around specific results by retrieving items before and after an anchor point (either specified or automatically found).
• Fetch: Retrieves the full details of selected observation IDs for comprehensive review.
• Supports filtering by type ("observations", "sessions", or "prompts") and observation type (bugfix, feature, decision, discovery, change).
• Allows date range filtering using dateStart and dateEnd.
• Offers ordering options like "date_desc," "date_asc," and "relevance."

Example prompts

• "Search for authentication issues in my-project and show me the last 20 results."
• "What was the timeline around observation ID 11131?"
• "Find bug fixes related to 'authentication' within the last week in project my-project."

Tips & gotchas

• Always follow the three-step workflow (search -> filter -> fetch). Skipping filtering can lead to unnecessary token usage.
• The skill is designed for retrieving information from previous sessions, not the current conversation context.
• When fetching details, use get_observations with multiple IDs in a single request whenever possible to optimize efficiency.