Knowledge

A Knowledge Article states useful context: facts, constraints, explanations, indexes and domains. It is the default Article type. Knowledge follows Karpathy's LLM-wiki pattern: immutable raw Sources, an agent-maintained Markdown wiki and its schema are ordinary files. Authored folder-index Articles are the summaries; there is no automatic recursive summarization.

Knowledge informs; it never grants. Retrieval may put a Knowledge Article in an Agent's context, but no link, similarity or Knowledge checkout can select a Task or grant a Tool.

Subjects in the starter Vault

The starter graph ships the project's own architecture and empty world subjects for a new owner to fill:

Subject	Contents
Architecture/	Shared design: the four Modules, with Harness/ and Shell/ implementation Articles. The Ontology lives here as Architecture/Harness/action-ontology.
Agents/Executive/Preferences/	Owner preferences. Empty in a fresh installation.
Games/, Projects/, Websites/, News & Research/	World subjects, initially just their folder Articles.
Media/	Media catalogue and delivery knowledge.

Placement rules: established preferences belong under the Executive; workstation procedures belong under Runbooks/Operations with explicit Skill dependencies; implementation explanations and dated incident evidence belong in Architecture/Shell; project knowledge belongs in Projects. The System namespace is reserved for the deterministic inventory publisher (below).

How Knowledge enters the Vault

Agents never edit accepted Knowledge directly. New evidence arrives as an immutable Source, is researched by Darwin, handed off through the physical Inbox, integrated by Alexandria and accepted through Review.

flowchart LR
    I["evidence/incoming/<br/>(drop folder)"] -->|source.added| D["Darwin · Learn"]
    W["web.fetch / explicit research"] --> D
    D -->|"source.handoff<br/>cited finding"| X["evidence/inbox/"]
    X -->|source.inbox| A["Alexandria · Ingest"]
    A -->|vault.propose| R{"Review"}
    R -->|approve| V["Vault + local Git audit"]
    R -->|reject| Z["proposal closed"]
    V --> G["retrieval index + graph"]

• Intake. Finished UTF-8 files in obsidience/evidence/incoming/ enter through filesystem events plus startup reconciliation (no polling; partial, hidden, binary, symlinked and oversized files are excluded; originals stay). A new durable Source emits source.added, which activates Darwin's Learn.
• Research. Darwin fully reads the Source, researches from primary sources and hands off one cited finding with source.handoff into obsidience/evidence/inbox/. Darwin owns research accuracy.
• Integration. source.inbox has one subscriber, Alexandria's Ingest, which turns the finding into the smallest coherent set of Article creates or updates with complete bodies. Alexandria owns integration.
• Missing knowledge. A small model must not guess around a gap. The Executive looks up public facts itself with web.search and web.fetch; separately queueable research is delegated as one bounded Question or Learn.

Maintenance

Alexandria's maintenance Tasks keep the wiki current. Each is a separate Task with its own Runbook (see Tasks):

Task	Does
Ingest	Integrates one Source Inbox handoff.
Curate	Runs one bounded inspection and may activate exact Merge or Link Tasks; or reviews one changed Hindsight mental-model page and stages at most three recommendations.
Merge	Consolidates one exact duplicate set, staging one canonical update before any archive and redirecting every inbound reference.
Link	Adds one meaningful relationship as a first-class Link review, with the exact wikilink, line and endpoint revisions.
Improve	Edits, categorizes, expands, retitles and splits.
Archive	Handles one deprecated or evidenced superseded Article, keeping content, Sources and history under _archived/.

Review

Review is the publication boundary for accepted Knowledge.

• One unresolved proposal per Article, pinned to its base revision; stale or overlapping proposals fail closed.
• Approval is atomic. A failed response is reconciled by reading, never by replaying the decision.
• The Harness decides whether a proposal is a Link or an Article change; the model and the UI never guess it.
• Approved changes commit to the Vault's own local Git repository (with no remote) when git_commit is on; SQLite remains the Review authority even without Git audit.
• Deterministically validated routine maintenance may auto-accept. Conflicts, destructive lifecycle changes, authority changes and high-risk effects always need Review.

Auto-curate is one owner-authored auto_curate boolean, inherited from the nearest explicit ancestor. It is permission, not truth: only scoped create/update Knowledge proposals may auto-approve. Identity, capability, authority, archives, conflicts, broken links, stale bases and memory-derived changes always keep Review. The starter Vault ships with Auto-curate off.

Retrieval and the Thinking Packet

Conversation and Tasks share one activation compiler. Each activation gives the model one visible Thinking Packet, each Article once, in order: the Agent identity; the Task and its acceptance (Tasks only); the immutable Objective; Tool Articles; paired Skills; applicable Runbooks (Tasks only); bindings and exclusions, including the newest Shell Scene; up to five Knowledge Articles; and the projected conversation history.

Knowledge comes from weighted reciprocal-rank fusion of a lexical and a dense lane over one accepted snapshot, with the Agent's Knowledge scope applied before each top-K. Staged, archived, temporary, expired, deprecated and retrieval: false Articles never enter. Hits are leads, not authority. The graph's thinking animation lights exactly this packet; see Graph UI.

System inventory

The System namespace is published by deterministic code (no model, Task or Review) at startup and on Refresh System: Compute, Devices, Drives, Network and one leaf per application, each citing its observed evidence. A failed collector keeps the last good Article and reports degraded coverage. Previous evidence is not proof of current state; current device claims require fresh observation.