How Trace works

Trace processes each source item through a durable sequence. Every derived fact remains connected to the Memory that supports it.

Source Connections

A Source Connection belongs to one Sterling Trace account and points to an external evidence source. It owns authorization state, import history and checkpoints. Identical evidence is deduplicated within that connection, never across customers.

Memories

A Memory is the preserved record of one source item: an email, document, image or bank-statement entry. Attachments can become linked Memories of their own, so the original evidence is never replaced by an interpretation.

Canonical Text

Extraction produces one complete text input for later stages. Email bodies remain part of it. Image-only evidence carries OCR provenance back to the original image and records which extraction attempt was accepted.

Enrichment

One enrichment pass derives four useful views from that Memory alone:

• Description — the complete, de-duplicated Markdown account.
• Gist — the shortest version that still preserves financially material context.
• Tags — controlled labels describing what the Memory is.
• Questions — retrieval hooks that make the Memory discoverable from related ideas.

Relevance and the graph

Trace records whether a Memory is accountant-relevant, then projects evidenced organisations and relationships into an owner-isolated graph. Verified public-source facts and the original Memory meet at a hard join before later accounting stages use them.

Processing state

Each expected artifact is pending, running, done, error or unprocessed. An error means a step that should have worked failed; unprocessed means Trace does not yet have that capability. The processing view and delivery logs expose those states without rewriting the evidence.