Synthetic memory
Memory that a system constructs from raw interactions — summaries, extracted knowledge cards, graph facts — rather than replaying the full transcript. In agent products, “synthetic” usually means consolidated, system-generated memory, not synthetic data. The risk is that the synthesis is lossy; the benefit is that the agent can carry context without carrying the whole session.
Read moreAgent memory
Persistent state an AI agent can write and retrieve across sessions: decisions, preferences, tasks, pitfalls, and context. Agent memory is what turns a stateless model into a stateful worker.
Read moreMemory layer
The infrastructure between the agent and the model that stores, indexes, retrieves, and consolidates memory. It is separate from the model and from the agent runtime, which is why it can be shared across IDEs and agents.
Read moreLong-term memory for LLMs
Memory that survives beyond a context window: stored externally, retrieved on demand, and updated over time. The model does not need to be retrained to use it.
Context engineering
Designing what enters the model’s context at each step: instructions, retrieved memory, tools, and state. Memory is one input to context engineering; it is not a replacement for it.
AI second brain
A personal or team knowledge layer that an agent can query across tools. In practice, it is a memory layer with a human-facing surface: notes, decisions, and links that persist beyond one chat.
Episodic memory
Memory of specific events and sessions: what happened, when, and with whom. It is useful for debugging and handoffs because it preserves the sequence of decisions.
Semantic memory
Memory of facts and concepts, independent of when they were learned. A semantic memory can be true across projects; an episodic memory is tied to a specific session.
Procedural memory
Memory of how to perform a task: workflows, skills, conventions, and tool sequences. In an agent, this is often encoded as skills or rules rather than as prose.
Memory consolidation
The process of turning raw events into durable, deduplicated knowledge. It can be rule-based or model-based; the important design question is what gets discarded and what stays auditable.
RAG vs memory
RAG retrieves from a relatively static corpus. Memory writes and updates from interactions, then retrieves. They can be combined, but a stateless RAG pipeline is not agent memory.
MCP memory
A memory server exposed through the Model Context Protocol, so any MCP-compatible agent can call init, recall, record, and lookup. MCP is the transport; the memory contract is separate.
Read moreLocal-first memory
Memory stored on the user’s machine as the primary copy, with cloud sync optional. The trade-off is that the user owns the data and the agent must work offline; the benefit is privacy and portability.
Read moreMemory benchmark
A reproducible evaluation of retrieval or answer quality. LongMemEval is one retrieval benchmark; Recall@5 measures whether the right memory is in the top five results, not whether the final answer is correct.
Read moreStateful agent
An agent that maintains state across steps or sessions. Memory is the persistence layer; the agent runtime decides when to read and write it.
Knowledge graph memory
Memory represented as typed entities and relationships, often with time bounds. It is useful for temporal and multi-hop questions that a flat vector index handles poorly.
Memory poisoning
Incorrect or adversarial memory that degrades future behavior. The defense is provenance, conflict detection, scoped writes, and the ability to audit and delete memory — not injecting malicious prompts into the web.
Context window
The model’s working token budget. Memory reduces the need to carry every prior message in the window; retrieval decides what enters it.