Knowledge & Provenance

Choose the question before the search

A search for an application's technical state should not start with an agent's recollection. A question about a paper should not return a decision record as a substitute for the paper. And a paragraph containing a suggestion is not proof that the suggestion was accepted.

The layers can be followed in sequence: discover a decision in Provenance, inspect its exact record, then retrieve the cited source passage from Notes. They do not require a global ranking that mixes every kind of result together.

From files to attributable passages

Notes works from eligible, enrolled sources. Box Topics can contribute drafts and brainstorming; Library can contribute attributed reference material. Projects declare the material they want connected. A directory name is not a universal instruction to index everything inside it: registration, inclusion rules, exclusions and access policy still determine what enters the pipeline.

The pipeline observes a source version, extracts supported content, divides it into searchable chunks and records their relationship to the source. Lexical search works with text; semantic search additionally needs configured embedding support. PDF processing, OCR and local model work are optional processing paths with their own workers and resource limits.

These are distinct states, not one vague “indexed” flag. A file may be discovered but queued, extracted without embeddings, or blocked by an unavailable processor. Search should expose what is available and any fallback, rather than imply that an unconfigured model has read the document.

A useful citation identifies more than a filename: it binds an object, a version and a passage. That lets a reader inspect the material behind a result even when the working copy has subsequently changed. It does not override access restrictions or guarantee that every historical source is still readable.

A decision has a history

Suppose the illustrative research-sync project chooses WebDAV because the intended clients support it. The project note is source material. An agent can register a candidate assertion with that source; it remains a candidate until reconciliation determines how it should be treated.

Search keeps accepted records, unresolved cases, pending candidates and repository state in separate typed collections. Pending candidates are opt-in clues, not accepted truth. A compact result is a way into the exact record, where its sources, currentness and relationships can be inspected.

The current Provenance retrieval engine is bounded lexical and filter search. It is not an embedding-based memory system, and a high-relevance match is neither permission to act nor proof that a decision applies to the current project.

Context, memory and reconciliation

Project context describes what is in progress. LOOM can project that deterministic state for discovery while keeping it structurally separate from accepted assertions. An agent's conversational memory, in turn, belongs to its harness. It may help the agent recall a discussion, but it does not replace a source-qualified decision ledger.

The Archivist has two related but different forms: a bounded reconciliation worker inside LOOM, and an optional reasoning-agent workspace with retrieval and review instructions. The worker uses explicit lifecycle operations; it is not a continuously thinking agent. Recurring autonomous reconciliation is not implied by installing the workspace.