Skip to main content

The memory system

Wren AI's memory layer is the second of five layers of context — the place where successful natural-language-to-SQL pairs, prior interactions, and user feedback accumulate so the agent gets better at querying your data the more it is used.

Memory is local. It is stored under .wren/memory/ in your project directory, indexed with LanceDB, and never leaves your machine unless you commit it to a shared Git repository.

What lives in memory​

  • NL-SQL pairs — ("How many customers placed more than one order?", "SELECT ... FROM customers ...") — stored when an agent runs a query that succeeds.
  • Pinned business questions — questions you mark as canonical examples; these are weighted higher during recall.
  • Schema and column descriptions — extracted from your MDL so retrieval can match both questions and structure.
  • Instructions — content from instructions.md is indexed alongside the rest so retrieval can surface relevant rules.

How recall works​

When the agent needs to answer a new question, it first asks the memory layer for the most relevant context:

  1. wren memory fetch --query "..." returns the most likely tables, columns, and relationships
  2. wren memory recall --query "..." returns the most similar past NL-SQL pairs
  3. The agent reads both and uses them as the grounding for its SQL generation

The more queries you and your team run, the more reliable recall becomes.

When to re-index​

Memory is automatically updated when you store a new pair (wren memory store), but the full LanceDB index is only rebuilt when you run wren memory index. Re-index after:

  • bulk-editing instructions.md
  • changing model descriptions in models/*/metadata.yml
  • importing a large batch of seed NL-SQL examples

See wren memory --help for the full set of commands.