Skip to content

Vector Memory Collections

The Vector Memory Collections are the lowest-level components of the Long Term Memory.
These are particular databases that store the content in the form of geometrical vectors.

A vector memory comes in the guise of a named collection of vectors and additional, optional metadata.
The latter can be used to filter the search in the database.
Each vector represents a memory. They are also called embeddings as they are the results of the text-to-vector conversion yielded by the embedder.

Such databases are particularly useful because they allow to fetch relevant documents based on the vector similarity between a query and the stored embeddings.

By default, Vector Memory Collections are created when the Cat is installed or after a complete memory swap.

Vector Memory Collections flow

flowchart LR
    subgraph CAT ["#128049;Cheshire Cat"]
        H["#129693;"]
        H1["#129693;"]
        direction LR
        subgraph LTM ["#128024;Long Term Memory"]
            direction TB
            C[(Episodic)];
            D[(Declarative)];
            P[(Procedural)]
        end
        H --> C
        H --> D
        H --> P
        C --> H1
        D --> H1
        P --> H1
    end
    E[First Installation] ----> H;
    F[Memory Swap] ----> H

Nodes with the 🪝 point the execution places where there is an available hook to customize the execution pipeline.