Memory

Every agent has its own long-term memory. It is what lets an agent pick a conversation back up after you close the chat.

What gets remembered

A memory is a short, self-contained statement with a type and an importance from 0 to 1:

Type	Example
fact	The project database is Postgres 16
decision	We decided to keep one SQLite file per workspace
preference	Val prefers short answers with code

Memories get created in three ways:

• The agent calls the remember tool when you tell it something worth keeping.
• After each turn, a background job asks the model to pull out durable facts from the exchange.
• You add them yourself in the Memory tab, with crew memory add, or through the API.

If a new memory is nearly identical to an existing one (cosine similarity of 0.9 or more), it is skipped.

How recall works

Search runs inside SQLite and combines two methods:

1. A full-text index (FTS5) ranked with BM25.
2. Embeddings from your embedding model (nomic-embed-text by default), stored as blobs and compared by cosine similarity.

The two rankings are merged with reciprocal-rank fusion, then adjusted a little for importance and age.

If no embedding model is available, Crew falls back to full-text search and fills in the embeddings later.

Memories with an importance of 0.8 or more are treated as core and always go into the agent's context.

Inspect and edit

crew memory Lena                     # list
crew memory Lena --search "stack"    # search
crew memory add Lena "Release is on Friday" --type decision

In the web UI, open an agent and go to Memory to search, add and delete.

TIP

Clearing a chat does not clear memory.