Knowledge, Notes & Collections — which do I use?

Blenau gives your AI agents three kinds of memory. They look similar from the outside — “somewhere the agent can store and retrieve things” — but they follow opposite principles, and using the wrong one is the most common way to make Blenau less useful, not more. This page is the map.

The three tiers at a glance

	Knowledge	Notes	Collections
What it holds	Curated understanding — docs you’ve written or crystallized	Disposable reminders — lists, to-dos, ideas	Structured records synced from a system (an Odoo ERP)
Who writes it	You, once you’ve understood something	You, from an impulse	Your source system, automatically
How long it lives	Indefinitely	Until you act on it	As long as the source keeps it
You retrieve it to…	Think	Act, then clear	Resolve a fuzzy question to an exact record id
Success looks like	It accumulates	It empties	It stays in sync
Backed by	Your GitHub repos (real commits)	Blenau only	Blenau (a mirror; truth stays in the source)

The one-line test:

Did you understand this and want to keep it? → Knowledge. Are you just offloading a task or an item? → Notes. Is it a record from another system you want to look up? → Collections.

Knowledge — your curated brain

The synthesized understanding that makes you and your agents better: how your auth works, why a decision was made, the runbook for a deploy. It earns its place because it passed through your comprehension, and it lives indefinitely, versioned as real commits in your own GitHub repos. You recall it to think.

• Use it for: documentation, decisions, runbooks, hard-won lessons, anything an agent should treat as reliable and cite.
• Don’t use it for: a shopping list, a one-off reminder, or a raw dump you haven’t digested — that dilutes what “the brain knows” with noise.
• Reach it with: search_knowledge, ingest_document, crystallize_session.

Get started with Knowledge →

Notes — your working memory

Quick, disposable things you want off your mind: a shopping list, “remind me to call the supplier”, an idea for the next video. Captured for free in a sentence, recalled complete and exact, and cleared as you act. It’s the agentic equivalent of a calendar — it frees your head, it doesn’t make you wiser.

• Use it for: lists, reminders, to-dos, fleeting ideas — anything meant to be done and then gone.
• Don’t use it for: anything you want to keep and reason over later — that’s Knowledge. Notes are never returned by search_knowledge, on purpose: asking your brain “what do I know about X?” must never surface a grocery item.
• Reach it with: remember, recall, mark_note_done, forget.

Learn about Notes →

Collections — your structured records

A live-ish mirror of records from a system you already run — products, customers, partners from an Odoo ERP. Blenau doesn’t own this data; it resolves a natural-language question (“the premium sateen sheets”) to the exact record’s id + metadata, so your agent can then fetch the volatile truth (stock, live price, order status) straight from the source with that id.

• Use it for: looking up business records by fuzzy description, filtering by typed fields, getting a stable id to act on.
• Don’t use it for: free-form notes or curated prose, and never as the source of live truth — a Collection is only as fresh as its last sync.
• Reach it with: list_collections, describe_collection, query_collection.

Learn about Collections →

Why keeping them separate matters

Each tier is optimized for its job, and mixing them quietly breaks the thing that makes it good:

• Put a grocery item in Knowledge and your agent starts citing it as something you know — the brain gets noisier and search gets worse.
• Treat Notes like a database and you lose the one guarantee it makes: recall returns the whole list, exact, so you can trust it.
• Treat a Collection like the source of truth and you’ll quote a stale price.

When in doubt, come back to the one-line test above. The rest of the docs go deep on each: Knowledge, Notes, Collections.

Want the model underneath this map — the primitives, and why grouping is done with paths instead of tags? See The mental model.