PLOWSHARE
Collect what matters, find it again and inspect the evidence behind a conclusion. Personal notes, immutable sources, memory and research reports each have a useful place in the system.
Use Personal for an evolving wiki, plans and source collection. The starter layout gives incoming captures, resources, plans, outgoing artifacts and archived material a home.
Save a thought, link, file or decision from a conversation with its origin and your reason for keeping it. Capture first; a saved link remains a link until its content has actually been acquired.
Read the sources, distill their essence and connect them to knowledge pages or plans. Preserve originals and label sources that remain pending or unavailable.
Retrieve relevant knowledge, turn a goal into a concrete next step, or draft a brief connected to its sources. Creating a plan does not start project work, and drafting a message does not send it.
Review progress and stalled decisions. Check links, stale claims, contradictions and orphan pages within a chosen scope. Preserve unresolved disagreements rather than smoothing them into certainty.
Capture, process inbox, recall, plan, express, review and maintain knowledge are explicitly granted procedures. Use them in Personal with the appropriate file access; their discoverability in another project does not expose your private content.
Retain pasted text or acquire a public URL. Each resource can have several immutable revisions, keeping older quotations attached to the exact text originally read.
The desktop Library has source and report collections with different document kinds. Browse by title or retained text, user tags, generated tags, category, author, saved date and format. Readable report drafts can be found before final publication.
Extraction, derived structure, embeddings, summaries and tagging have separate readiness states. Retained text can be readable while a model stage is incomplete. Lexical filtering can work without semantic vectors; vector retrieval needs compatible prepared projections.
User tags belong to the resource and survive revisions. Automatic tags are model-generated, revision-specific suggestions. They help navigation and remain separate from the labels you set and the evidence you quote.
A document question runs a bounded workflow over its corpus and consumes model work. Inspect its cited text and explicit gaps. Reading a saved source or filtering the Library does not rerun research.
Evidence records contain a quote, its source revision and exact retained-text coordinates. Replacing a source creates a new revision instead of moving old citations.
| Keep | When you need |
|---|---|
| Workspace note | An editable plan, wiki page or synthesis under file permissions. |
| Information source | Versioned reference text with an immutable retained revision. |
| Evidence quotation | Exact support tied to the text you actually read. |
| Research report | Structured findings, supporting inputs, citations and producer lineage. |
| Memory or digest | A lesson, decision trail or bounded summary connected to its formation history. |
Private drafts do not become shared because a bot discusses them. Report access depends on consumed inputs, including uncited material. Exclusion removes a revision from discovery; withdrawal blocks its content and dependent outputs. Those are different lifecycle decisions.
The bundled deep research orchestration combines source acquisition, analytical workers, adversarial assessment and report assembly in an inspectable procedure.
Set the topic, scope, objectives and output expectations. Choose a calling assistant granted the workflow and check the serving models, search provider and budgets.
Reviewed queries lead to retained source revisions and evidence acquisition. A fetch receipt does not establish that extraction succeeded or that the source was used.
Analytical and adversarial workers assess support and counter-evidence. Findings can hold, weaken, be refuted or remain unchecked; unavailable assessment stays a visible gap.
Read the report alongside its orchestration audit and inputs. Feedback can lead to a new revision. Readability, finalisation and sharing are separate states.
Search availability, source failures, incomplete expansion and model quality affect the outcome. Source preparation and delegated research have separate allowances; inspect actual usage instead of treating the caller’s limit as the total cost.
Deep research and report inspection ↗Memories retain useful lessons with provenance and lifecycle. Conversation retrieval and digests help recover how earlier decisions developed.
Memory recall finds relevant items; reading fetches an actual retained item. Proposals, review and curation have their own contracts. Historical or invalidated knowledge should remain labelled when it appears in advice.
Memory digests ↗Compaction bounds future model context while preserving recorded entries under their retention policy. Digests link back to source spans. Precise claims and quotations still need the original material.
Conversation retrieval ↗