The language behind the boundary.

Context bleed

Also: context bleeding, context carryover

The unwanted transfer of facts, instructions, tone, memories, permissions or assumptions from one context into another where they no longer apply or should not be available.

Context contamination

Also: context pollution

The presence of material that distorts later reasoning or generation. Contamination may be accidental, operational or adversarial and does not necessarily involve disclosure.

Context engineering

The discipline of deciding which instructions, messages, memories, documents, tools and state should be available to a model at each step of a workflow.

Context leakage

Also: cross-context information leak

A confidentiality or isolation failure in which contextual information is disclosed to another user, tenant, agent, project or workflow without the required authorisation.

Context pruning

Also: context compaction

A context-engineering technique that drops, compresses or archives stale, duplicated, irrelevant or low-confidence information to protect focus and reduce cost.

Context rot

Also: long-context degradation

A practitioner term for non-uniform loss of accuracy or coherence as inputs become longer, noisier or more internally competitive, even when the context remains within the advertised window.

Context window

Also: input window

The bounded sequence of tokens supplied to a model when it generates a response, including instructions, messages, retrieved material and tool results.

Irrelevant-context distraction

Also: contextual distraction, contextual interference

A measurable sensitivity in which distractor passages, examples or facts change a model's reasoning path or answer despite not being needed for the task.

Lost in the middle

Also: middle-position bias

A positional-bias phenomenon in which language models retrieve or use information less reliably when it appears in the middle rather than near the beginning or end of a long context.

Memory isolation

Also: memory segregation

Architectural controls that partition persistent state by user, tenant, project, agent or purpose and prevent unintended cross-boundary retrieval or mutation.

Memory poisoning

Also: persistent-memory poisoning

An attack or integrity failure in which false, malicious or misleading content is written into agent memory and later retrieved as trusted context.

Memory retrieval

The process that ranks and injects stored memories into a current context, often using recency, semantic similarity, importance or explicit policy.

Persistent memory

Also: long-term agent memory

A product or agent capability that stores facts, preferences, summaries or events outside the immediate context window and can retrieve them in future sessions.

Prompt injection

Also: direct prompt injection, indirect prompt injection

An attack in which instructions supplied directly or embedded in external material attempt to manipulate model behaviour, data access or tool use.

Provenance

Also: data lineage

Metadata that identifies a context item's source, owner, permissions, timestamp and transformation history so systems and reviewers can assess its authority and suitability.

RAG poisoning

Also: retrieval poisoning

An integrity attack in which misleading content or adversarial instructions are placed in a retrievable corpus so they influence later model outputs at query time.

Retrieval leakage

Also: cross-tenant retrieval

A failure to enforce permissions before retrieval, allowing protected chunks, embeddings or records to enter model context and potentially be disclosed.

Retrieval-augmented generation

Also: RAG

A system pattern in which a search or retrieval component supplies selected knowledge to a generative model so its response can be grounded in information outside model parameters.

Sensitive information disclosure

Also: data disclosure

The exposure of personal, proprietary, regulated or security-relevant information beyond its intended audience, whether through generation, retrieval, logs or tool calls.

Tenant isolation

Also: multi-tenant isolation

Technical separation of data, memory, caches, indexes, credentials and execution state belonging to different organisations or customers sharing a service.

Trust boundary

An architectural boundary at which identity, authorisation, provenance and data handling must be re-evaluated rather than implicitly inherited.

Vector store

Also: vector database

Storage and indexing infrastructure used to retrieve semantically similar items. In multi-tenant AI systems it must preserve access-control metadata at chunk level.