Glossary

The terms that come up most when talking about AI memory, in plain language.

Memory layer

The infrastructure piece that gives an AI system persistent memory: it ingests knowledge, structures it and retrieves it on demand. It's what stops your agent from starting from zero every session.

MCP (Model Context Protocol)

A standard protocol for giving an AI assistant access to external tools and data. MemoryFirst exposes an MCP server (stdio, HTTP, SSE) to remember and retrieve memory like any other tool.

RAG (Retrieval-Augmented Generation)

A pattern where the model, before answering, retrieves relevant passages and uses them as context. Answer quality depends above all on retrieval quality.

Fact graph

Representing knowledge as entities connected by relationships, not just loose text. It lets you retrieve by connection, not only by similarity.

Embeddings

A numeric representation of a text's meaning. Dense search compares embeddings to find semantically similar passages even when they share no words.

BM25

A lexical search algorithm that scores by exact term match. It shines where meaning isn't enough: codes, proper nouns, case numbers.

RRF (Reciprocal Rank Fusion)

A method to combine several result lists by scoring each document by its rank in each list, instead of by scores that aren't comparable across systems.

Reranking (cross-encoder)

Fine reordering of candidates by looking at the query and each passage together (not separately, like embeddings). More expensive, so it's applied only to the top of the pool.

Citation / Provenance

The verifiable origin of a claim: the paragraph, the minute, the exact message. An AI that asserts without citing isn't memory — it's a hallucination with a good reputation.

Hallucination

When a model asserts something plausible but false or unsupported. Citations and provenance-aware retrieval are the main defense.

Context window

How much text a model can hold “in view” in one call. It's working memory, not storage: it's wiped when the call ends. That's why you need a persistent memory layer.

AI-first

Building the product on a layer that accumulates context instead of discarding it, rather than bolting a chatbot on top. The model is a commodity; the memory is yours.

GDPR

The EU data protection regulation. For a memory layer it means data residency, a data processing agreement (DPA), rights of access/export/deletion and no training on customer data.

Self-hosting

Running the memory layer on your own infrastructure, with no phone-home. What enters your instance stays in your instance.