Cortex

• A recall answer names the prior sessions, ISAs and decisions that match, ranked, each with its date and path, and states what was decided or done there rather than only the topic; when nothing matches, it says so and names the query and the sources searched.
• An archive search returns a deduplicated table of matching notes (title, domain, status, tags), or an explicit "no matches" with a suggested next step; retrieve returns the top matching notes ranked, with compressed summaries.
• A captured learning tells the problem, assumption, reality, journey with its dead ends, solution and takeaway, filed under LEARNING/SYSTEM/YYYY-MM/ or LEARNING/ALGORITHM/YYYY-MM/; a recent-history digest lists the files it read and says nothing they do not.
• All archive operations (add, harvest, develop, ingest) classify each note into one of the six canonical KNOWLEDGE note types and add typed cross-links (related: frontmatter) between notes to keep the graph connected: every new note carries the eight envelope fields and 2-4 typed related: links, each naming a slug that exists in the archive.
• Every result comes from the command's loaded deterministic tools, and any verification that was unavailable is named in the result as unavailable, never reported as passed.

• Never proceeds on a partial load: a truncated read is not a successful load, and truncation warnings count anywhere in the output, not only at the END CONTEXTLOAD marker.
• A failed search is never taken to establish absence.
• No archive write drops the loaded schema, link, or approval requirements.
• The result never misses a material relevant record, attributes content to the wrong source, or proposes an unauthorized write.
• No personal configuration is kept inside this public skill.

These two run from their intent files, not a ContextLoad.ts bundle; both work on the LEARNING store (LIFEOS/MEMORY/LEARNING/) and saved research, never the KNOWLEDGE archive.

/cortex <args> and legacy /knowledge <args> route identically. Arguments that do not name a subcommand are archive searches.

Before executing any Cortex command, load its bundle from the Intent Routing table (or full, when the command does not fit a listed bundle or relevance is uncertain) with the deterministic loader: bun "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/LIFEOS/TOOLS/ContextLoad.ts" Cortex load BUNDLE

It returns the selected command instructions and mandatory archive constraints together, with source hashes. Follow every returned requirement. When the loader fails, read References/CompleteGuide.md and the relevant command section directly. Paths in the preserved guide are relative to the skill root. Already-loaded, unchanged sections need no second read; after compaction or a source change, load the relevant bundle again.

For bundle reads, set the shell tool output budget to 30000 tokens. With a code-mode wrapper, also set its output budget to 30000.

Use list to discover sections and bundles, and search "query" to search the preserved guide. When a search finds nothing, inspect the index, broaden the query, or load full.

The distill command reads its optional configuration from LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Cortex/DistillConfig.json, as its loaded instructions describe. Keep personal configuration outside this public skill.

• None recorded at skill level yet. Command-level gotchas live in References/CompleteGuide.md § Gotchas and load with every bundle.

Example 1: Recall prior work

User: "recall the invoice for the Texas trip"
→ recall bundle
→ Ranked prior sessions and ISAs with dates and paths, and what each one settled

Example 2: Add a person

User: "add people Jane Doe"
→ add bundle
→ A People note with the eight envelope fields and 2-4 typed related: links to notes that exist

Example 3: Search the archive

User: "what do we know about prompt injection?"
→ Routes to search — 3-pass (lexical + frontmatter + wikilink) over MEMORY/KNOWLEDGE/
→ Returns table of matching notes with domain, status, tags

Example 4: Ingest a source

User: "/knowledge ingest https://example.com/article"
→ Fetches the source, classifies entity type, creates primary note with typed related: links
→ Ripple pass proposes updates to existing related notes; user approves; MOCs regenerated

Example 5: Status check

User: "knowledge status"
→ Runs KnowledgeHarvester.ts status
→ Shows domain note counts, orphan wikilinks, stale seedlings, time since last harvest