---
title: "What is vector memory?"
description: "Vector memory defined: storing text as embeddings and retrieving by similarity, what it does well, and what it cannot track by itself."
canonical: https://past.dev/glossary/vector-memory
last-updated: 2026-09-02
---
# What is vector memory?

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

Vector memory stores text as embeddings and retrieves by semantic similarity; it finds related content but does not by itself track time or truth. Text is embedded into vectors, stored in an index, and retrieved by nearest-neighbor search against an embedded query. This works well when relevance means resemblance. It has no native notion of when a fact was true, whether it was later superseded, or which source asserted it.

## How vector memory works

On write, text is split into chunks and each chunk is embedded into a vector that encodes its meaning. On read, the query is embedded the same way and the index returns the nearest stored vectors. The matching chunks are then placed into the model's [context window](/glossary/context-window). Chunk size is a tuning decision: small chunks match precisely but lose surrounding context, while large chunks keep context but blur what matched.

Because matching happens in embedding space, a query can find text that shares meaning without sharing words. A search for "pricing complaint" can surface a chunk that says the renewal quote felt too high.

## What vector memory is good at

- **Paraphrase matching.** Questions rarely reuse the wording of the source text; embeddings match them anyway.
- **Unstructured content.** It works directly on prose, with no schema or modeling step required.
- **Low setup cost.** Embed, index, query: the pipeline is short and well supported by [vector databases](/vector-database-for-rag).

## What it does not track

Vector memory ranks by resemblance. It has no native representation of time, supersession, or source. Suppose a February 3 note says the project budget is $40,000 and a May 20 note says the budget moved to $60,000. Both chunks sit near the query "project budget", and similarity scores say nothing about which value is current. Sorting by insertion order does not repair this either, because records arrive in ingestion order rather than event order.

The same applies to conflicts and attribution: when stored chunks disagree, retrieval returns both without a status value, and a chunk does not record which source asserted it or whether that source was later removed. Teams address this by adding metadata filters, timestamps, and reranking on top, or by using a memory system that models facts explicitly. The full comparison is at [vector databases vs memory](/vector-database-vs-memory), and the [benchmark methodology](/benchmarks/methodology) explains how these differences are measured.

## Related concepts

- **[Hybrid retrieval](/glossary/hybrid-retrieval)**: similarity search combined with keyword and structured filters.
- **[Memory graph](/glossary/memory-graph)**: entity-and-relationship storage, the structured alternative.
- **[Memory staleness](/glossary/memory-staleness)**: the failure mode when superseded chunks keep ranking high.
- **[Context rot](/glossary/context-rot)**: what happens when too many retrieved chunks fill the window.

## Frequently asked questions

### What is vector memory in AI agents?

It is memory implemented as an embedding index: text is stored as vectors and retrieved by semantic similarity to the query. It is the most common starting point for giving an agent recall over past text.

### What are the limitations of vector memory?

Similarity search has no built-in concept of time, supersession, or source. When stored chunks disagree, it returns the closest matches and leaves the conflict for the model to resolve.

### Is a vector database enough for agent memory?

For paraphrase-style lookup over documents it can be. For facts that change over time, exact identifiers, or answers that need dates and sources, teams add structure on top or use a memory system that tracks facts explicitly.

## Related

- [What is a memory graph?](https://past.dev/glossary/memory-graph)
- [What is hybrid retrieval?](https://past.dev/glossary/hybrid-retrieval)
- [What is memory staleness?](https://past.dev/glossary/memory-staleness)
- [Facts that change over time](https://past.dev/guides/facts-that-change-over-time)
- [Memory API quickstart](https://past.dev/docs/memory-api/quickstart)