knowledges
the memory of the house
local grounding is the default: an answer grounded in the curated topic bank beats model memory and beats a fresh web fetch. search first, load only the one topic that matches, and repeat the search whenever task scope changes. vector results are discovery, never authority: open the selected topic file before acting on it.
searches before assumingpoints, never copiesstays indexed
Local knowledge grounding: always search the local knowledge index before loading concept, library, platform, procedure, playbook, domain, pattern, protocol, language, format, or tool topics.
methodology
- look up by exact topic name first, then keyword search, then regex, then semantic search for conceptual or differently-worded tasks.
- search by type with the bank's search script:
--keyword "<query>" --type <type>. - read only the selected topic file (plus its references when needed); never bulk-load the bank.
- when the request describes an outcome but no procedure, search with
--type procedure, choose the narrowest matching topic, and follow it. - check index freshness with the bank's
statusscript; if stale, rebuild it withindexbefore trusting its results.
contents
- trainedtested local search behavior and limits
- topicsthe typed topic bank: read the one matched topic file under concepts, libraries, platforms, procedures, playbooks, patterns, domains, languages, protocols, formats, or tools
- scriptssearch and index implementation: retrieval modes, chunking limits, dedup keys, scoring weights, manifest staleness
- assetsthe local vector store backing semantic search and its manifest
- namingnaming law for skill and topic names before creating or renaming one