zeekay
aad2b6998c
The single binary brain that compounds knowledge across every channel, every agent, every runtime. Drop markdown into ~/.hanzo/workspace/. Edges auto-extract (zero LLM). Facts queryable via MCP. One SQLite file at ~/.hanzo/brain/brain.db. Same file readable by every Hanzo runtime — TS, Python, Rust, Go. Packages - brain meta-pack registering memory + graph-links + recipes - memory pluggable BrainStore (SQLite default; FTS5 hybrid search) - graph-links zero-LLM typed-edge extractor (6 edge types) - recipes-brain YAML daily-life automations (email flagship) Sister runtimes - TS canonical ships in @hanzo/bot (this monorepo is the source) - Python hanzo-memory (graph_links + recipes ports) - Rust hanzoai/mcp::brain crate - Go hanzoai/bot-go (single static binary) Canonical paths - ~/.hanzo/brain/brain.db the brain - ~/.hanzo/workspace/ markdown source — auto-ingested - ~/.hanzo/recipes/ user-authored YAML recipes - ~/.hanzo/config.toml per-machine config - ~/.hanzo/cache/ embedding cache, tool-output cache - ~/.hanzo/logs/ structured logs A brain.db written by ANY runtime is byte-identical to one written by every other. Same schema, same FTS5 setup, same pages / edges / facts shape. Pluggable backends — same BrainStore contract everywhere: sqlite (default) · qdrant · meilisearch · zapdb · replicate · vfs · postgres.
@hanzo/bot-recipes-brain
YAML ingest recipes for the Hanzo Brain. Daily-life automations — email today, calendar / Slack / X / Notion next.
A recipe defines: auth + cron + ingest + classify + draft + enqueue + on_swipe hooks. The bot runs it on a schedule, drops drafts into a queue, the Hanzo app shows the swipe UI. Every swipe is a training signal.
Flagship: email
hanzo-bot brain run-recipe email
hanzo-bot brain drafts list
hanzo-bot brain drafts approve <id> # sends via gmail
hanzo-bot brain drafts reject <id> # logs as decline-fact for training
See recipes/email.yaml for the spec. Auth is OIDC via hanzo.id (single sign-on across all Hanzo services). Cron defaults to every 30 minutes — adjust per inbox volume.
How a recipe runs
cron fires
→ ingest fetch new items from backend (gmail, calendar, …)
→ classify cheap model tags each item (needs_reply, priority, …)
→ draft main model writes a response using brain context:
• brain.recall(from_email) — facts about sender
• brain.search(thread_id) — prior thread
• brain.facts({entity, since:30d}) — recent activity
• brain.style(sender:me, recent:200) — your tone
→ enqueue drop into drafts.email queue (TTL 7d)
→ notify push to the Hanzo iOS / Android / macOS app
Swipe → train loop
Every swipe in the app feeds back into the brain:
- swipe right (send) → send via gmail, upsert fact
{subject:from_email, predicate:"replied"}, add(received, sent)pair to style examples - swipe left (reject) → delete draft, upsert fact
{predicate:"declined_reply"}, down-weight similar drafts - swipe up (snooze 24h) → reschedule
- tap edit → send → upsert fact
{predicate:"replied"}, add the edited version to style examples (highest-quality training pair)
After ~200 swipes, the draft quality compounds — the brain learns your tone from your own actions, not from a tone-tuning UI.
Adding a new recipe
Drop a <name>.yaml into recipes/. The loader walks the directory; the scheduler registers anything with a cron: field.
Minimal shape:
recipe: notion
version: 1
backend: notion
auth:
provider: hanzo.id
cron: "0 */2 * * *"
ingest:
source: workspace-pages
since: 12h
classify:
model: zen-2-haiku
schema: { needs_reply: bool, priority: one_of [P0,P1,P2,P3] }
draft:
when: needs_reply == true
model: zen-2
context:
- brain.recall(page_id)
- brain.search(title)
enqueue:
queue: drafts.notion
ttl: 7d
API
import { listRecipes, loadRecipe } from "@hanzo/bot-recipes-brain";
await listRecipes(); // ["email", "notion", …]
const recipe = await loadRecipe("email");
Plugin contract
import register from "@hanzo/bot-recipes-brain";
const { recipes } = await register(api, ["email", "notion"]);
// → registers each recipe as a cron job in the bot scheduler
YAML driver
Bot already depends on the yaml package (eemeli/yaml v2). No new deps.