About the Agent Memory Glossary
A reference layer for one specific surface of the agentic-AI domain — the memory architecture of LLM agents. Built and maintained by the AgentsBooks team.
Why a memory-only glossary
By 2026 the question for anyone designing an agent fleet has moved from “which model?” to “how does it remember?”. Working memory, episodic memory, semantic memory, procedural memory, RAG retrievers, vector stores, knowledge graphs, context graphs, MemGPT-style paged memory, Mem0-style fact extraction, MemoryBank-style forgetting curves — the vocabulary fragmented quickly, and the same word now means different things in different stacks.
This glossary is the single look-it-up surface for that vocabulary. Every entry is one canonical primary-source citation, dated. Each one resolves cleanly into a copy-pasteable definition for an agent prompt, a design doc, or an internal review.
The framing
“We introduce RAG models which combine pre-trained parametric and non-parametric memory for language generation.”
“Context engineering is the practice of curating, ordering, compressing, and refreshing the context an LLM sees on each call.”
“A novel architecture and interaction patterns enabling believable simulations of human behaviour — agents that remember, reflect, and plan over time.”
Audience
ML engineers wiring up retrieval pipelines, knowledge architects deciding between vector and graph stores, agent system designers building multi-step workflows, and operators evaluating agent runtimes. This site assumes you can read an architecture diagram; it does not assume you know the difference between IVF and HNSW — that is the point of looking it up.
The eight primitives
The vocabulary AgentsBooks uses for everything: Identity · Brain · Heart · Memory · Control · Knowledge · Friends · Shares. Two of the eight — Memory and Knowledge — are the subject of this glossary. See Anatomy of a Firm for the full set.
Family of properties
This is one of a small, growing family of authority properties published by AgentsBooks. The companion sites:
- Agentic Glossary — deep, narrative entries for ~31 flagship terms across the whole agentic-AI space.
- Agentic Glossary — Quick Reference — the broader companion to the deep glossary; 88 agentic-AI terms across protocols, primitives, and operations.
This site is the third member of the glossary family, scoped exclusively to memory architecture — what differentiates an agent from a stateless chatbot.
Voice and discipline
- Operator-grade, plural first-person, citation-anchored.
- No hyperbole without a number; no editorialised “X is the best” without a method footnote.
- Vendors are referenced as compatible substrates, not as adversaries. Pinecone, Weaviate, Qdrant, Milvus, pgvector, Anthropic, OpenAI, Cohere, Microsoft Research, Neo4j, the W3C — every term that touches a vendor surface cites that vendor's own primary documentation.
- When a term has a longer entry in the deep glossary, this site links there instead of repeating the long form.