🧠 Tool: @recall

The @recall tool is the retrieval engine for the SQLite long-term memory system, allowing the agent to pull past context dynamically based on search queries or categories.

Technical Implementation

The agent invokes @recall to execute a SELECT query against the project's memories table. Unlike a global prompt injection, this tool allows for selective retrieval, keeping the prompt window clean while providing the necessary context.

The LLM can provide:

• query: A string for simple fuzzy matching.
• category: A filter to narrow down results (e.g., searching only in Architecture).
• limit: Controls how many facts are returned (defaults to the 5 most recent).

Usage Examples

Example 1: Retrieval for Implementation Alignment

User

How did we implement auth in this project?

Zed Agent

Searching project history for authentication patterns...

Tool Output

@recall({ query: "authentication", category: "Architecture" }) >> Found: "Base auth uses JWT with crates/core/src/auth.rs."

Example 2: Debugging Recurring Issues

User

I'm getting a strange generic error in the `axum` handler. Have we seen this before?

Zed Agent

Searching project history for archived issue resolutions...

Tool Output

@recall({ query: "axum error", category: "Issues" }) >> Found: "Axum generic error fixed by adding `Send` bounds to the state type."

Workflow Impact

• Deep Search: Instantly filters through thousands of historical facts to find the one relevant to your current task.
• Consistency: Ensures the agent provides answers that align with past decisions rather than hallucinating new patterns.
• Error Reduction: Proactively retrieves "Lessons Learned" from the Issues category to prevent repeating past mistakes.