ke — Knowledge Engine

One CLI tool, one skill. Binary: ke (~/.bun/bin/ke; fallback bun ~/repos/ke/bin/ke-tool.ts). KB at ~/vault/ke/ (local sqlite-vec index, embeddings via Xenova/all-MiniLM-L6-v2 — no server, no API key). ke with no args prints full usage.

Verbs

Verb	Command	Use
search	`ke search "<q>" [--limit N] [--tag T] [--diverse]`	Semantic vector search → ranked `[score%] title → path`. Look up prior decisions/fixes/patterns before acting.
research	`ke research "<q>" [--depth N]`	RECALL + fill ≤N gaps web→ingest. `--depth 0` = recall-only (no fetch/write). Each depth = one web→ingest pass on the top remaining gap; gaps return as a value, you decide whether to go deeper.
update	`ke update "<path>" --content "..."` / `echo … \| ke add "<topic/title>"`	Persist a durable learning. Queue at session end, after a fix, after research.
learn	`ke learn "<fact\|@file\|url>" [--topic T]`	Ingest caller-supplied material directly (atomizes + indexes, `src:learn`). No crawl.
select	`ke select [--tag T] [--topic T] [--regex P] [--after D] [--before D] [--format yaml]`	Exact-match filter (AND logic). Machine-readable with `--format yaml`.
pending	(see "background jobs" below)	Inspect background `ke research` jobs the prompt-hook spawned.
others	`query`, `compile`, `list`, `maintain`, `stats`, `audit`, `dispute`/`support`/`resolve`	Run `ke` with no args for usage.

Recall before deriving

Search BEFORE re-deriving anything plausibly studied before (experiments, infra gotchas, library pins). Hooks in hooks/ already auto-inject precedent on prompt/stop in wired projects — don't re-search what a hook just injected.