Manages a typed knowledge archive of linked notes and recalls prior sessions and conversations.
npx skills add https://github.com/ruban-s/DevOS --skill cortex-ruban-s
Please help me install this Agent Skill. Skill: Cortex Source: https://github.com/ruban-s/DevOS/tree/main/skills/Cortex Command: npx skills add https://github.com/ruban-s/DevOS --skill cortex-ruban-s
related: edges (supports, contradicts, extends, etc.) so the collection behaves as a graph, not a folder. - Ingest with Ripple: Absorbs a URL or file, drafts a note with baked-in edges, then propagates updates and contradiction flags to neighboring notes. - Session Recall: A deterministic Bun CLI fans out over five local stores (work registry, session names, work dirs, spec bodies, conversation transcripts) to find prior work by theme or phrases like "yesterday" or "last week". - Use Case: Ask "what do we know about prompt injection?" to get tabulated archive hits, or run /cortex recall yesterday markdown to resume yesterday's work with ranked snippets and file paths. ## Quick Start Ask the agent to search the knowledge archive for a topic, for example: what do we know about prompt injection?High-intent search queries and answers about installing and using this skill.
Run the search flow with a query to sweep the archive lexically, by frontmatter title and tags, and by wikilinks using ripgrep. Results are tabulated with domain, status, and tags. For scored, compressed context, use the retrieve command backed by MemoryRetriever.ts.
Use the recall command with a topic or date phrasing like "yesterday" or "last week". The ContextSearch CLI scans work.json, session names, work directories, ISA spec bodies, and conversation JSONL transcripts, scoring token overlap weighted by recency.
The archive accepts exactly four shapes: People, Companies, Ideas, and Research. Every note must carry 2-4 typed related: edges such as supports, contradicts, or extends, plus body wikilinks, validated against the schema in _schema.md.
Yes, the ingest flow fetches a URL or reads a file, classifies it into a domain, drafts a note with typed edges, then proposes a ripple plan updating neighboring notes and flagging contradictions before applying changes and reindexing.
Recall scoring is token-overlap, not substring matching, so queries sharing no non-stopword tokens with stored content miss. Date filters like "yesterday" also bound results; JSONL dating uses the first user-message timestamp rather than file mtime.
It is not suited for broad published-content search across blogs and feeds or one-shot URL ingestion through other harvester pipelines. Content that fails the lookup test—things you would never retrieve by name—belongs in work or learning trees instead.
rijnbConvert raw notes, VTT, EML, and HTML files into Markdown with provenance metadata.
fractaryCreate and maintain bidirectional spec-to-GitHub issue links via comments and descriptions.
davduncManage a persistent Knowledge Archive with typed cross-linked notes.
oakossManage Turso edge-hosted SQLite databases with embedded replicas and vector search.