Memory is durable state. It is not whatever text remains in the prompt.
An Agent OS separates short-term working memory from durable memory so that agents can continue across turns, devices, workers, and sessions.
Current operational state:
- active goal
- active constraints
- current floor
- running workers
- open approvals
- latest artifacts
Working memory is optimized for action.
What happened:
- traces
- utterances
- tool calls
- worker attempts
- retries
- failures
- user corrections
Episodic memory is optimized for audit and debugging.
What the system believes:
- facts
- preferences
- relationships
- domain knowledge
- decisions
Semantic memory requires source and confidence.
How the system should act:
- skills
- playbooks
- eval cases
- failure mitigations
- style guides
- workflow rules
Procedural memory is optimized for repeatable behavior.
type Belief = {
id: string;
claim: string;
source: string;
confidence: number;
scope?: string;
createdAt: number;
updatedAt: number;
expiresAt?: number;
};Before writing memory:
- Identify the source.
- Assign confidence.
- Classify freshness.
- Check whether it conflicts with existing memory.
- Prefer appending an auditable correction over silent overwrite.
Memory should be consolidated after:
- repeated user correction
- failed live test
- repeated worker failure
- new artifact accepted by user
- production incident
- eval gap
Some memory should expire:
- dynamic market facts
- pricing
- schedules
- API limits
- temporary preferences
- one-off constraints
Forgetting is part of correctness.
If memory cannot name its source, it is not memory. It is model residue.