Glossary

The product terms below describe stored records and the operations that use them. Human-memory analogies and vocabulary from other projects do not imply identical APIs, retention policies, or access boundaries.

Library terminology

Term Meaning in Neo4j Agent Memory

Term	Meaning in Neo4j Agent Memory
Short-term memory	Persisted conversations and messages. The name describes conversational use, not automatic expiry or in-process storage. On Bolt, sequence uses edges including `:FIRST_MESSAGE` and `:NEXT_MESSAGE`.
Long-term memory	Reusable declarative knowledge: entities and, on supported backend surfaces, preferences and facts. See Backend capabilities.
Reasoning memory	Recorded task steps, tool calls, observations, and outcomes. It contains what the application or integration records, not a model’s private internal reasoning. Note: "observation" here is a reasoning step’s recorded observation, stored alongside the step’s thought and action and often derived from a tool result. It is a different sense from the context-compression "Observation" row below.
POLE+O	Person, Object, Location, Event, and Organization: the default broad entity categories. See POLE+O.
`:Entity`	A label for managed entity nodes. On Bolt it accompanies type/subtype labels and identifying properties. It does not by itself establish ownership or authorization.
Adoption	Attaching the managed entity label and required properties to a compatible existing graph. See Adopt an existing graph.
Backend	The storage interface used by a client: direct Neo4j through Bolt, or hosted NAMS. Shared concepts do not imply identical operations, parameters, or result shapes.
NAMS	Neo4j Agent Memory Service, the hosted REST service. Its service API and a particular SDK release are distinct versioned surfaces.
Workspace	The NAMS tenancy boundary. A workspace key is bound to it; an admin key selects a workspace according to the authentication contract.
Conversation or session	A grouping of conversation/task records. Identifier terminology and creation differ by backend. A conversation ID is not itself an access-control boundary.
User identifier	An application user identifier accepted by specific operations. It is not a universal filter on every record or method. See Scope by operation.
Bridge protocol	The separate method-oriented transport used by conformance reference services. It is not the hosted NAMS REST API. Use the documented REST transport for hosted examples.
Provider string	A model identifier resolved by the Python provider factory, such as `openai/text-embedding-3-small`. It configures client-side providers; NAMS manages its own extraction and embeddings.
Agent Skill (preview)	A portable procedure package derived from a scoped set of NAMS memory records. See Agent Skills.
Skill distillation	The server-side process that selects memory, synthesizes a procedure, checks grounding, and packages a skill. It is separate from executing the procedure.
Provenance grounding	Links from a generated claim or step to its source records, such as `:GROUNDED_IN` edges. These aid review; they do not prove that the source is true or the generated interpretation is correct.
Attestation	A signature used to check a published skill package’s integrity and signer. It does not validate the factual correctness of its contents. See Skills API.
Observation (context tier)	A NAMS-generated summary of a window of conversation messages, part of the three-tier context model. Both SDKs read it on NAMS: TypeScript `getObservations()` and Python `short_term.get_observations()`. This is distinct from the "observation" in the Reasoning memory row above, which belongs to one reasoning step, not a compressed message summary. The Python MCP server’s `memory_get_observations` tool returns a third kind: facts and decisions extracted by its local observational memory, available on Bolt.
Reflection	A NAMS-generated, higher-level synthesis derived from observations once accumulated context crosses a compression threshold, returned by TypeScript `getReflections()` and Python `short_term.get_reflections()` on NAMS. The Python MCP server’s local observational memory also generates its own session reflections. See Understanding the three-tier context model.

Short-term memory

Persisted conversations and messages. The name describes conversational use, not automatic expiry or in-process storage. On Bolt, sequence uses edges including :FIRST_MESSAGE and :NEXT_MESSAGE.

Long-term memory

Reusable declarative knowledge: entities and, on supported backend surfaces, preferences and facts. See Backend capabilities.

Reasoning memory

Recorded task steps, tool calls, observations, and outcomes. It contains what the application or integration records, not a model’s private internal reasoning. Note: "observation" here is a reasoning step’s recorded observation, stored alongside the step’s thought and action and often derived from a tool result. It is a different sense from the context-compression "Observation" row below.

POLE+O

Person, Object, Location, Event, and Organization: the default broad entity categories. See POLE+O.

:Entity

A label for managed entity nodes. On Bolt it accompanies type/subtype labels and identifying properties. It does not by itself establish ownership or authorization.

Adoption

Attaching the managed entity label and required properties to a compatible existing graph. See Adopt an existing graph.

Backend

The storage interface used by a client: direct Neo4j through Bolt, or hosted NAMS. Shared concepts do not imply identical operations, parameters, or result shapes.

NAMS

Neo4j Agent Memory Service, the hosted REST service. Its service API and a particular SDK release are distinct versioned surfaces.

Workspace

The NAMS tenancy boundary. A workspace key is bound to it; an admin key selects a workspace according to the authentication contract.

Conversation or session

A grouping of conversation/task records. Identifier terminology and creation differ by backend. A conversation ID is not itself an access-control boundary.

User identifier

An application user identifier accepted by specific operations. It is not a universal filter on every record or method. See Scope by operation.

Bridge protocol

The separate method-oriented transport used by conformance reference services. It is not the hosted NAMS REST API. Use the documented REST transport for hosted examples.

Provider string

A model identifier resolved by the Python provider factory, such as openai/text-embedding-3-small. It configures client-side providers; NAMS manages its own extraction and embeddings.

Agent Skill (preview)

A portable procedure package derived from a scoped set of NAMS memory records. See Agent Skills.

Skill distillation

The server-side process that selects memory, synthesizes a procedure, checks grounding, and packages a skill. It is separate from executing the procedure.

Provenance grounding

Links from a generated claim or step to its source records, such as :GROUNDED_IN edges. These aid review; they do not prove that the source is true or the generated interpretation is correct.

Attestation

A signature used to check a published skill package’s integrity and signer. It does not validate the factual correctness of its contents. See Skills API.

Observation (context tier)

A NAMS-generated summary of a window of conversation messages, part of the three-tier context model. Both SDKs read it on NAMS: TypeScript getObservations() and Python short_term.get_observations(). This is distinct from the "observation" in the Reasoning memory row above, which belongs to one reasoning step, not a compressed message summary. The Python MCP server’s memory_get_observations tool returns a third kind: facts and decisions extracted by its local observational memory, available on Bolt.

Reflection

A NAMS-generated, higher-level synthesis derived from observations once accumulated context crosses a compression threshold, returned by TypeScript getReflections() and Python short_term.get_reflections() on NAMS. The Python MCP server’s local observational memory also generates its own session reflections. See Understanding the three-tier context model.

Memory analogies

These are approximate conceptual mappings, not a compatibility table for other products.

Term	Closest concept here	Limit of the analogy
Context graph	Connected records across the memory layers	The useful context for one call is a selected subset, not necessarily the whole graph.
Semantic memory	Reusable declarative knowledge in long-term memory	Semantic search is a retrieval technique and can also search messages or traces.
Episodic memory	Records of events, including conversations and task traces	The product does not assign this term exclusively to short-term or reasoning memory.
Procedural memory	Reusable knowledge about performing a task	A recorded trace describes one execution; deriving a reusable procedure or skill is a separate step.
Working memory	The context assembled for the current model call	It is not a separately persisted fourth layer. Stored records influence the call only when the application retrieves and supplies them.
Core memory	Application-selected durable information	No product-wide mapping to a specific prompt region or automatic retrieval behavior is implied.
Memification	A loose term for memory ingestion and transformation	Use the concrete operation name here: extraction, resolution, embedding, storage, or retrieval.

Term

Closest concept here

Limit of the analogy

Context graph

Connected records across the memory layers

The useful context for one call is a selected subset, not necessarily the whole graph.

Semantic memory

Reusable declarative knowledge in long-term memory

Semantic search is a retrieval technique and can also search messages or traces.

Episodic memory

Records of events, including conversations and task traces

The product does not assign this term exclusively to short-term or reasoning memory.

Procedural memory

Reusable knowledge about performing a task

A recorded trace describes one execution; deriving a reusable procedure or skill is a separate step.

Working memory

The context assembled for the current model call

It is not a separately persisted fourth layer. Stored records influence the call only when the application retrieves and supplies them.

Core memory

Application-selected durable information

No product-wide mapping to a specific prompt region or automatic retrieval behavior is implied.

Memification

A loose term for memory ingestion and transformation

Use the concrete operation name here: extraction, resolution, embedding, storage, or retrieval.

Glossary

Library terminology

Memory analogies

See also