Glossary: AI memory in terms.

AI agent

An AI agent is a language model that does not just answer but carries out steps on its own: reading files, running commands, changing code, calling tools. Examples are Claude Code, Codex or Hermes. For a single request an agent often performs dozens of steps, and each one costs tokens.

AI memory

An AI memory stores knowledge across sessions: decisions, project facts, settings, past mistakes. A language model on its own forgets everything after each session that is not in its training data. More in the guide AI memory.

Second brain

A second brain is an external memory for notes, ideas and knowledge. With AI it becomes a system that does not just file content but links it by itself and brings it back at the right moment. More in Second brain AI.

Context window

The context window is the amount of text a language model sees at once for a request: system prompt, context files, conversation history, tool results. Whatever is not in it, the model does not know at that moment. The fuller it is, the more expensive and often less precise the answer. More in the guide Context window.

Token

A token is the unit language models compute in, usually part of a word. Billing is based on the tokens that go into and come out of the model. Everything in the context window counts on every call, including content that contributes nothing to the current question. What Hebby costs and saves in tokens: Token efficiency.

Context file (CLAUDE.md, AGENTS.md)

A context file is a text file an agent reads automatically in every session, such as CLAUDE.md for Claude Code, AGENTS.md for Codex, GEMINI.md or MEMORY.md. It is the simplest form of memory, but it is sent in full with every call and grows over time. Why lean context files are better: Lean context.

Prompt caching

With prompt caching the model provider remembers the unchanged beginning of a request and charges less for it on the next call. That lowers the cost of repeated content such as context files, but does not make it free.

MCP (Model Context Protocol)

The Model Context Protocol is an open standard that lets AI applications connect external tools and data sources. A service that speaks MCP works in many clients, such as Claude Code, Codex or other agents, without a separate adaptation for each client.

MCP memory server

An MCP memory server gives an agent a memory over MCP: the agent can store and query knowledge. Hebby is connected this way and additionally brings matching memories in advance. More in MCP memory server.

Hook

A hook is a small script an agent runs automatically at fixed points, for example before a request goes to the model. Through hooks a memory can put memories into the context before the agent asks for them.

RAG (retrieval-augmented generation)

RAG means that before answering, a system searches a collection for matching passages and passes them to the language model. This lets the model answer with knowledge that is not in its training data. The quality depends on how well the search finds the right passages.

Embedding and vector search

An embedding turns a text into a series of numbers (a vector) that represents its meaning. Texts with similar meaning end up close to each other as vectors. Vector search uses this to find content with matching meaning quickly in large collections, even without the same words.

Reranker

A reranker sorts the results of a first, fast search a second time, thoroughly, by scoring question and result together. That improves the results but costs time and compute on every request. More in the guide Reranker.

Graph memory

A graph memory stores knowledge as a network of nodes and connections. That way the memory knows what belongs together, for example which decision belongs to which project. Hebby uses a graph memory of its own design. More in Graph memory.

Hebbian learning

Hebbian learning goes back to the psychologist Donald Hebb: what is active together connects more strongly. In Hebby, connections between memories grow stronger when they are used together. More in Hebbian learning.

Consolidation

In memory research, consolidation means that the brain sorts and strengthens new content afterwards and links it with what is already there, above all during sleep. An AI memory can mirror this: it merges duplicates, reorganizes and retires what is outdated.

Anticipatory recall

Anticipatory recall means the memory brings matching memories into the context before the agent searches for them. Research steps and follow-up questions drop away, and the agent knows known pitfalls before the first attempt.

AI knowledge management

AI knowledge management means collecting and organizing knowledge in teams or companies so that AI systems can use it, with permissions per person and per agent. More in the guide AI knowledge management.