Plain-English definitions of the terms you run into when giving an AI agent memory. No jargon left unexplained. Skim the whole list or jump straight to a term.
Agent memory is external storage an AI agent reads from and writes to so it can retain information across turns and sessions. The underlying language model keeps nothing on its own, so the agent does the remembering by saving to storage and reading it back. See our beginner guide on agent memory explained.
The context window is the block of text a language model can see in a single request. It is cleared after each request, which is exactly why an agent cannot rely on it for lasting memory. A bigger context window holds more of one conversation, but it still empties out and cannot reach back to earlier sessions.
An embedding is a numerical vector that represents the meaning of a piece of text. Texts with similar meaning have similar embeddings, which is what makes semantic search possible. Generating embeddings requires an embedding model, one of the moving parts you avoid when your memory is key-value rather than semantic.
Key-value memory is memory stored as named values that you retrieve by an exact key. You store French under the key user_language, then read that key back later. It is the simplest form of agent memory and needs no embeddings or vector database. AgentRAM is a key-value memory API.
MCP is an open standard that lets AI clients call external tools and services in a consistent way. A memory API exposed over MCP appears as a set of tools inside MCP-compatible clients such as Claude Desktop. AgentRAM ships an official MCP server on npm, so an agent can store and recall memories as native tool calls.
A namespace is a scope that isolates or groups memories. It keeps one agent's data separate from another's, or, when shared, lets several agents read and write a common pool. Shared namespaces are how multi-agent systems coordinate memory without passing everything through prompts.
Persistent memory is memory that survives after a session ends, so an agent can recall information in a completely new conversation. It is the opposite of the context window, which resets constantly. See our fuller guide on persistent memory for AI agents.
RAG is a pattern where an agent retrieves relevant text from a store and adds it to the prompt before the model answers. It is often built on semantic search over a vector database. Agent memory and RAG overlap: reading a memory into the prompt before responding is a simple form of retrieval augmentation.
Semantic memory is memory you search by meaning rather than by exact key. It is used when an agent must find the passages most relevant to a query across freeform text, and it relies on embeddings and a vector database. If you can only describe what you want by its meaning, not name its key, you need semantic memory.
Semantic search finds stored text by meaning rather than exact keywords. It compares the embedding of a query against the embeddings of stored text and returns the closest matches. It is powerful for large, unstructured collections, and it is the main reason a memory system would need a vector database.
Structured memory is memory where you know what you stored and roughly how you will ask for it. Preferences, state, and results are all structured memory. Because you can name the key, it needs no semantic search, which means no embeddings and no vector database. Learn where this line sits in our guide on agent memory without a vector database.
Temporal memory tracks facts along with the time they were true, so an agent can reason about how information changed. If a user moves cities, temporal memory records the change rather than storing both as current. It is a specialized need that dedicated temporal-graph tools handle.
TTL is an optional expiry on a stored memory, so it is automatically removed after a set period. It is useful for information that should not persist forever, such as a temporary session detail. In AgentRAM, a memory can be given a TTL in days when it is stored.
A vector database stores embeddings and retrieves them by similarity. It is the engine behind semantic search over large bodies of unstructured text. It is powerful but adds real setup and cost, which is unnecessary when your memory is key-value. Our guide explains when you do and don't need one.
AgentRAM is key-value agent memory over HTTP. No embeddings, no vector database, no setup. Store your first memory in about a minute.
Get your API key100 free operations. No credit card.
© 2026 AgentRAM. All rights reserved.