Each term is a short, citable definition for an engineer or an AI answer
engine. The coined terms link back into the Pillar 1 guide set.
agent-memory Agent memory
A durable layer that stores scoped, source-linked facts and makes them available to agents across sessions, tools, and model calls.
working-memory Working memory
Temporary state the agent uses for the current task, conversation, or tool run.
short-term-memory Short-term memory
Conversation or thread state that persists briefly but is not necessarily durable across tools or projects.
long-term-memory Long-term memory
Stored context intended to survive beyond the current session and be recalled later under scope and policy controls.
episodic-memory Episodic memory
Memory of events or interactions, such as a prior decision, support exchange, or coding session.
semantic-memory Semantic memory
Memory of facts about products, users, teams, systems, or the domain the agent works in.
procedural-memory Procedural memory
Memory of preferred processes, workflows, commands, or operating rules an agent should reuse.
source-memory Source memory
Memory tied to documents, tickets, pages, transcripts, or artifacts that can be cited and inspected.
context-window Context window
The temporary working space a model reads for one answer or bounded conversation.
context-engineering Context engineering
The practice of choosing and shaping the information an agent receives so it can complete a task reliably.
retrieval-augmented-generation Retrieval-augmented generation
A pattern where source material is retrieved and passed to the model so the answer can be grounded in evidence.
memory-write-path Memory write path
The process that decides what should be remembered and records source, scope, type, and lifecycle metadata.
memory-read-path Memory read path
The process that retrieves relevant memories, ranks them, and returns evidence to the agent or user.
memory-scope Memory scope
The boundary that decides who or what can read a memory, such as tenant, workspace, library, user, or client.
tenant-isolation Tenant isolation
A safety property where one tenant's memory cannot be retrieved or exposed inside another tenant's answers.
source-artifact Source artifact
The canonical stored object, such as a document version or transcript segment, from which memory artifacts are derived.
provenance Provenance
The origin and lineage of a memory, including the source system, artifact, version, and derivation path.
citation-span Citation span
The specific part of a source that supports an answer, claim, or recalled memory.
retrieval-trace Retrieval trace
A record of the query, candidates, ranks, returned memories, citations, and answer path for one retrieval event.
inspectable-memory Inspectable memory
Agent memory that exposes what was remembered, why it was retrieved, where it came from, and whether it is current.
black-box-memory Black-box memory
Memory that influences answers without exposing source lineage, retrieval reasons, scope, or correction paths.
supersession Supersession
A replacement relationship where a newer source or memory marks an older one as no longer current.
superseded-memory Superseded memory
A remembered fact that remains inspectable as history but should not be treated as the current answer.
current-state-flag Current-state flag
Metadata that marks whether a source artifact or memory should be treated as the current version.
decay Decay
A memory lifecycle behavior where older or weaker signals become less likely to retrieve over time.
amnesia Amnesia
A forgetting behavior that removes temporary conversational carryover without necessarily deleting source artifacts.
purge Purge
A deletion behavior that removes or disables a source and the derived memory artifacts controlled by the system.
drift Drift
A state where a source or memory changes without a clean replacement relationship, making recall suspect.
provable-forgetting Provable forgetting
The ability to show that a memory and its derived artifacts were removed, disabled, or superseded under policy.
delete-handle Delete handle
A durable identifier that lets a product target the source and derived artifacts affected by a deletion request.
tombstone Tombstone
A record that marks a memory or source as removed or disabled so stale references do not continue serving.
derived-artifact Derived artifact
A summary, embedding, tag, extracted entity, or citation object produced from a source artifact.
source-lineage Source lineage
The chain from original source to stored artifact, derived memory, retrieval event, and cited answer.
embedding Embedding
A numerical representation of text or data that helps retrieval systems find semantically similar items.
vector-index Vector index
A search index that retrieves items by similarity between embeddings.
hybrid-search Hybrid search
Retrieval that combines semantic vector search with lexical or full-text search.
reranker Re-ranker
A model or scoring step that reorders retrieved candidates before the final answer uses them.
golden-set Golden set
A hand-reviewed query and answer set used to evaluate retrieval or answer quality reproducibly.
mcp-memory-server MCP memory server
A Model Context Protocol server that exposes memory tools to compatible agent clients.
context-pack Context pack
A scoped bundle of retrieved evidence prepared for an agent to use in a task.
tool-call Tool call
A structured request from an agent to an external tool, API, or MCP server.
agent-client Agent client
The tool or runtime where the agent operates, such as Claude Code, Cursor, or another MCP-compatible client.
memory-portability Memory portability
The ability to use, export, or migrate memory across agent clients and frameworks without lock-in.
no-training-default No-training default
A product posture where customer memory is used for recall and product operation, not to train foundation models by default.
permission-aware-retrieval Permission-aware retrieval
Retrieval that applies access scope before memories can be returned to an agent or answer.
retrieval-log Retrieval log
A durable event record of what a memory system searched, retrieved, ranked, and returned.
memory-conflict Memory conflict
A state where two or more memories disagree and the system needs inspection, supersession, or correction.
audit-trail Audit trail
A record of memory actions, including writes, retrievals, corrections, deletion requests, and lifecycle changes.
roadmap-claim Roadmap claim
A capability described as planned or shipping next rather than presented as available now.