---
title: "What is a memory graph?"
description: "A memory graph stores agent knowledge as entities and dated relationships. Definition, contrast with vector stores, and the temporal variant."
canonical: https://past.dev/glossary/memory-graph
last-updated: 2026-09-02
---
# What is a memory graph?

Source: https://past.dev/glossary/memory-graph

A memory graph stores an agent's knowledge as entities and dated relationships rather than as isolated text chunks. Each node is a person, organization, or subject; each edge is a fact connecting them, carrying the time it applies to. Because the structure is explicit, a query can follow relationships and filter by date rather than matching similar text alone.

## How a memory graph works

A memory graph is built from ingested text rather than modeled by hand. Extraction turns each record into facts, [entity resolution](/glossary/entity-resolution) links the subjects of those facts to existing nodes, and each fact becomes a dated edge between nodes.

A plain knowledge graph stores what is currently true, while a [temporal knowledge graph](/glossary/temporal-knowledge-graph) also records when each fact became valid, when it stopped being valid, and which fact replaced it. Memory graphs for agents are usually temporal, because the subjects they track change: owners rotate, prices move, policies get revised. Construction techniques and query patterns are covered in [knowledge graphs for LLMs](/knowledge-graph-for-llm).

## Memory graph vs vector store

| Property | Vector store | Memory graph |
| --- | --- | --- |
| What is stored | Text chunks with embeddings | Entities and dated facts |
| How retrieval works | Similarity to the query | Relationships, dates, and filters |
| What an update does | A new chunk is added beside the old one | The old fact's validity window closes; the new fact links to it |

The two are complementary. [Vector memory](/glossary/vector-memory) finds text that resembles the question; a graph answers questions about structure and time, such as who owned an account in March. Hybrid systems use both signals.

## A memory graph in practice

An agent ingests three records about Meridian Health. A January 8 email names Dana Whitfield as the account owner. A March 3 note records that the contract renews on October 1. A June 11 transcript records Dana handing the account to Tomas Rivera.

The graph stores one Meridian Health node, one node per person, and dated edges: an ownership edge for Dana valid from January 8 to June 11, an ownership edge for Tomas valid from June 11, and a renewal fact recorded on March 3. A July query for the current owner returns Tomas with the June 11 handoff as evidence, and a query about March returns Dana. past.dev builds this structure from ingested text automatically: entities are resolved, and each fact carries event time, a validity window, and a supersession link; see the [quickstart](/docs/memory-api/quickstart).

## Related concepts

- **[Temporal knowledge graph](/glossary/temporal-knowledge-graph)**: the time-aware graph model most memory graphs use.
- **[Vector memory](/glossary/vector-memory)**: similarity-based storage, the usual alternative or companion.
- **[Fact supersession](/glossary/fact-supersession)**: how a changed fact replaces its predecessor inside the graph.
- **[Entity resolution](/glossary/entity-resolution)**: how references across records land on the same node.

## Frequently asked questions

### What is a memory graph in AI?

It is a way of storing an agent's long-term memory as a graph: entities as nodes, facts as dated relationships between them. Retrieval follows structure and dates instead of relying only on text similarity.

### How is a memory graph different from a vector database?

A vector database stores text chunks and retrieves by semantic similarity. A memory graph stores entities and dated facts, so it can answer who is connected to what, as of which date, and what changed.

### What is the difference between a memory graph and a knowledge graph?

A memory graph is a knowledge graph used as an agent's memory. It is usually built automatically from ingested text and extended with timestamps so facts can be superseded rather than overwritten.

## Related

- [What is a temporal knowledge graph?](https://past.dev/glossary/temporal-knowledge-graph)
- [What is vector memory?](https://past.dev/glossary/vector-memory)
- [What is entity resolution?](https://past.dev/glossary/entity-resolution)
- [Facts that change over time](https://past.dev/guides/facts-that-change-over-time)
- [Benchmark methodology](https://past.dev/benchmarks/methodology)