What your agent remembers

Promise 1 is the agent knows you, your brands, and what’s on your plate. Promise 3 is it carries forward what you’ve decided together. Both come down to memory — and memory is where agent systems usually cheat, by stuffing the whole chat history back in and calling it remembering. WilsonOS does something more deliberate.

Five libraries, one precedence

An agent’s knowledge comes from five mutually exclusive libraries. It loads them in order, and an earlier library wins on conflict.

#	Library	What is in it	Who writes it
1	You	Your role, preferences, schedule, working style, hard rules — “never X”, “always Y”	You, and Mia when you state a preference
2	Brand	Each brand’s voice, visuals, customer, products, strategy, lessons	The brand’s agent, as brand state evolves
3	Capability	The skills and flows themselves — the toolkit	Whoever authors or improves a skill
4	The system	How the platform itself works — its architecture, tools, operations	The platform lane
5	Memory	The long tail: facts, decisions, events, corrections, working notes	Each agent, continuously, through sanctioned writers

The precedence is the point. If a skill says “lead with the keyword” and the brand’s voice says “lead with the reader’s problem”, the brand wins. If a remembered fact contradicts something you have stated as a rule about yourself, you win. The agent never has to guess which authority to follow.

What is not memory

If you can look it up from the current state — a skill’s instructions, a brand’s voice document, the app’s routes — it is not memory; it is one of libraries 2 to 4. Memory (library 5) is for what is not otherwise discoverable: a preference you expressed once, a decision made in conversation, a lesson from a result.

Two layers: what is stored, and what is loaded

Storage and loading are separate questions, and confusing them is the classic mistake.

Stored: brand-shared knowledge (libraries 2–4) lives as documents, read-only to running agents. Private and conversational memory (library 5) lives in a database, where every row carries whose it is, which agent it belongs to, and — for conversation memory — which conversation. Search indexes cover both.

Loaded: at session start an agent gets a lean, scoped bundle — see Session start. Everything else is pulled in by the skill or flow that needs it. No agent loads a full brand stack or the architecture “just in case”.

Private stays private

Memory has scopes, and scope is checked by trusted code on every read and write — not by a prompt instruction.

• Your memory with an agent — durable preferences, facts, decisions, a rolling summary — is readable by you, through that agent.
• A conversation’s memory — its events and its own summary — is readable by that conversation’s participants, and is never merged across conversations.
• Brand-shared knowledge is readable by everyone with access to the brand.

A private note never becomes shared brand knowledge by an agent’s judgment. The only path from private memory into a brand’s shared lessons is promotion: Mia proposes an entry in the weekly review, you approve it, and the shared entry records where it came from. Nothing an automation writes lands in a person’s private memory.

Memory is written on purpose

Agents record decisions and facts as they happen, through sanctioned writers, during and at the end of a task — not by anyone re-reading the chat log afterwards. Nightly consolidation distils recent events into a fresh summary and a ranked digest of facts; a weekly review promotes what deserves to be shared. Those processes can only distil what was written. A decision made in conversation and never recorded is lost, which is why “write it down now” is a rule and not a nicety.

Finding things

Retrieval is deliberately simple: a navigation manifest per brand, links between brand documents, a backlinks index, keyword search over the whole corpus, and each skill declaring up front the documents it needs. Dense vector search is deferred until this approach demonstrably fails on real work — it has real cost and real maintenance, and at the current corpus size the payoff is unproven. See Recall.

live The five libraries and their precedence, scoped private and conversation memory checked in trusted code, the sanctioned writers, nightly consolidation, and the promotion lane are live. direction Migration of older, unscoped notes into the scoped store is under the owner’s review, row by row.

Related

Session startExactly what loads when an agent wakes.

RecallHow an agent finds what it needs without loading everything.

SharpeningNightly consolidation, the weekly review, and promotion.

Access & permissionsThe scopes memory is checked against.