Fundamentals

Memory Model

Neutron AI organizes memory by Nucleus and Scope. This keeps integrations predictable across products, agents, workspaces, and users.

Core Terms

Term	Meaning
`nucleusId`	Your top-level customer, workspace, organization, or application partition. `tenantId` is a legacy alias.
`scopeId`	A narrower memory boundary such as `user:<id>`, `workspace:<id>`, `project:<id>`, or `agent:<id>`.
Memory Cell	A typed, durable fact, preference, procedure, entity, relationship, decision, outcome, context, or lesson that may improve future agent work.
Entity	A stable domain-neutral identity with attributes, confidence, validity, version, and provenance.
Relationship	A first-class directed or bidirectional connection with attributes, graph dimension, evidence, confidence, and validity.
Context	A bounded set of eligible memory, graph state, decisions, consequences, outcomes, procedures, contradictions, and reviewed learning returned for one task.
Privacy class	A label that helps your integration decide what can be recalled, cached, shown, or omitted.
Tombstone	A deletion marker that prevents removed memory from reappearing.

Recommended Scope Design

Choose scopes that match how your product grants access.

Product need	Example scope
Personal assistant memory	`user:<user-id>`
Team shared knowledge	`workspace:<workspace-id>`
Project-specific agent context	`project:<project-id>`
Support account history	`account:<account-id>`
Short-lived workflow memory	`session:<session-id>`

You can request multiple scopes for an agent task. Keep the list explicit and only include scopes the current user or agent is allowed to access.

Memory Types

Use memory type labels consistently so recall quality improves over time.

Type	Use for
`preference`	Stable user or workspace preferences.
`semantic`	Lasting domain facts and concepts.
`episodic`	Events and experiences tied to a time or workflow.
`entity` and `relationship`	Memories that support stable graph identities and connections.
`procedural` and `skill`	Standing methods, rules, and verified workflows.
`decision` and `outcome`	What was considered, selected, predicted, and later observed.
`contextual`, `working_session`, and `long_term`	Task, session, and durable context with explicit lifecycle intent.
`outcome`	What happened after a task, decision, or tool run.

Lifecycle

flowchart LR
  A["remember"] --> B["connect entities"]
  B --> C["build context"]
  C --> D["compare decisions"]
  D --> E["observe outcomes"]
  E --> F["review learning"]
  F --> A
  A --> G["tombstone"]

Use forget and graph tombstones for deletion, correction, user privacy controls, and stale or unsafe knowledge. Deletions should be treated as final in current integration logic. Historical versions remain available only where the authorised temporal and audit model requires them.

Continue with Knowledge and decision intelligence for temporal graph, Context Engine, decision, outcome, simulation, and learning workflows.

Architecture

Compression