Architecture + Implementation

[TanGentleman]

As I continue working on the memory system from different angles, here is what I've learned:

• Superior document processing is informed by the use case downstream
• Original sources, like a PDF file or audio segment, exist somewhere, and we possess a pointer. We should not duplicate all source documents.
• Temporal data is informative.

All Convex tables:

• sources
• journals
• entities
• relations
• knowledge
• embeddings
• metadata
• operations

Architecture

1. Storage Layer: A set of sources and their parts. When used in downstream tasks, these are cited.
2. Distilled Layer: A set of structured journals that always cite sources. These should be validated for correctness before being added to the index.
3. Knowledge Layer: A mapping of associations between entities and their relations. These are grounded in journals or human intervention, not sources.

Journals should be:

1. Easy for the user to read, understand, and update
2. Grounded by source documents. We are agnostic to the structure and contents, as long as there is a valid reference to a source and its parts.