Glossary · 40+ terms

AI Memory Glossary

Plain-language definitions of AI memory terms — from working memory and episodic memory to embeddings, RAG, consolidation and vector stores — the vocabulary used across AI Memory Works.

Core

Core concepts

AI memory

External storage and retrieval that lets LLM agents remember information across turns and sessions — distinct from the context window, RAG and fine-tuning. Full guide →

Agent memory

Memory systems built for autonomous AI agents — write, store, retrieve, consolidate and forget. Synonymous with AI memory in this knowledge base.

Long-term memory

Information persisted across sessions in external stores (vector DB, graph, KV). Guide →

Short-term memory

Working memory in the context window — current task and recent turns. Guide →

Working memory

Same as short-term memory for agents — what the model sees this turn. Not computer hardware RAM.

Types

Memory types

Episodic memory

Specific past interactions the agent can reference — event logs with timestamps. Guide →

Semantic memory

Stable facts and learned knowledge — preferences, account details, general knowledge. Guide →

Procedural memory

Skills, tools and how-to procedures — how to do things, not just facts. Guide →

Parametric memory

Knowledge stored in model weights (fine-tuning) — not updatable at runtime. Guide →

Non-parametric memory

External memory stores — vectors, graphs, KV — updatable without retraining. Most agent memory is non-parametric.

Shared memory

Common store read and written by multiple agents. Guide →

Architecture

Architecture terms

Memory retrieval

Searching and ranking stored memories to inject into context. Guide →

Memory consolidation

Merging and promoting short-term memories to durable long-term storage. Guide →

Virtual context

Treating the context window like RAM and paging memories in/out. Guide →

MemGPT

Research architecture for virtual-context memory paging. Implemented in Letta. Guide →

Context rot

Accuracy degradation when stuffing too much into a long context window. Guide →

Memory as a tool

Exposing memory operations as callable agent tools. Guide →

Infrastructure

Infrastructure terms

Embeddings

Vector representations of text for semantic similarity search. Guide →

Vector database

Database optimized for similarity search on embeddings — common LTM backend. Guide →

Knowledge graph

Graph store for facts, relationships and temporal validity. Guide →

RAG

Retrieval-augmented generation — retrieving from a fixed document corpus. Not the same as agent memory. Guide →

Graphiti

Temporal knowledge-graph engine used by Zep for agent memory.

Frameworks

Frameworks and benchmarks

Engram

Weaviate’s vector-native memory layer. Guide →

LangMem

Long-term memory for LangGraph. Guide →

LOCOMO

Benchmark for long-conversation memory recall. Guide →

LongMemEval

Benchmark for cross-session memory recall. Guide →

Disambiguation

Commonly confused terms

Limited memory AI

Not computer hardware RAM. In AI agent context, refers to agents with bounded working memory (context window) that use external stores for long-term recall. Human vs AI memory →

LSTM vs agent memory

LSTM is a neural network architecture for sequences — not the cognitive-science memory taxonomy for agents. Agent memory = external vector/graph/KV stores.

Context memory

Working memory in the context window — what the agent sees this turn. Distinct from long-term external memory.

FAQ

Frequently asked questions

What is AI agent memory?

External storage and retrieval that lets LLM agents remember information across turns and sessions — write, store, retrieve, consolidate and forget. See glossary: AI memory and the complete guide.

Episodic vs semantic memory — what's the difference?

Episodic = specific past events ("user asked about refund Tuesday"). Semantic = stable facts ("user prefers email"). See episodic and semantic glossary entries.

What does MemGPT mean?

A research architecture treating context as RAM with memory paging in/out. Letta implements MemGPT. See glossary: MemGPT.

What is a vector store?

A database that indexes embedding vectors for similarity search — the typical backend for semantic long-term memory. See glossary: vector database and vector databases for memory.

RAG vs memory — are they the same?

No. RAG retrieves from fixed documents; agent memory is dynamic and personal. See memory vs RAG.

What is limited memory AI?

Not hardware RAM. In agent context it means bounded working memory (context window) plus external long-term stores. See glossary disambiguation.

What is context rot?

Accuracy degradation when too much history is stuffed into a long context window. External memory with selective retrieval mitigates it. See context rot.

What is LOCOMO?

A public benchmark testing recall from very long single conversations. See LOCOMO benchmark.