Vector Database vs Memory Layer: Which Layer Are You Actually Missing?
A vector database stores and searches. A memory layer decides what is true, who a fact refers to, when it expired, and who may see it. When similarity search is enough, and when it is not.
TL;DR
A vector database and a memory layer are not competing products, they are different layers of the same stack. The vector database is storage and similarity search. A memory layer decides what is true, resolves who and what a fact refers to, records when it became true, and enforces who may see it. Most teams evaluating "vector database versus memory" are really asking whether similarity search alone is enough, and the answer depends on whether their questions are lookups or reconstructions.
The comparison people think they are making
Teams usually arrive at this question after a specific failure. The agent worked in the demo, then gave two different answers to the same question a week apart, and now somebody is asking whether the vector database was the wrong choice.
It usually was not. A vector database does exactly what it claims: it stores embeddings and returns the nearest ones. What failed is that nothing in the stack was responsible for deciding which of the returned passages was current, which entity they referred to, or whether the person asking was allowed to see all of them.
That responsibility has to live somewhere. If it does not live in a layer, it lives in whoever writes the query, which is why these systems degrade as they grow.
What each layer is responsible for
- Embedding model. Turns text into vectors. Choice affects retrieval quality more than most teams expect and it is worth benchmarking on your own corpus.
- Vector database. Stores vectors and returns nearest neighbours fast, at scale, with filtering. Pinecone, Weaviate, Chroma, pgvector and others compete here on latency, cost and operational burden.
- Retrieval orchestration. Chunking, query rewriting, reranking, top-k selection. This is where most RAG tuning effort goes.
- Memory layer. Resolves entities across sources so one customer is one customer, records valid time separately from ingestion time so superseded facts can be demoted, attaches permissions at the level of an individual fact, and keeps citations so any claim traces to its origin. Sentra sits here.
- Agent harness. Decides what to do with the answer.
A memory layer typically uses a vector database underneath. Asking which to pick is like asking whether to choose a filesystem or a database.
When similarity search alone is enough
Genuinely often, and it is worth saying plainly.
If your corpus is mostly static, one team owns it, questions are self-contained, everyone who can query it may see all of it, and nobody needs to know when a fact changed, then a well-tuned retrieval pipeline over a vector database will serve you for a long time. Adding a memory layer to that situation buys complexity you do not need.
It stops being enough under five conditions: identity fragments across systems, facts get superseded, answers require traversing a chain rather than reading a passage, different readers are cleared differently, or context volume is substituting for precision. On that last point, a peer-reviewed study across six model configurations found an answer placed mid-list among twenty retrieved documents scored 53.8 percent, worse than the 56.1 percent the same model scored with no documents at all.
| Capability | Sentra memory layer | Vector database plus RAG |
|---|---|---|
| Nearest-neighbour search | Uses one underneath | Yes, this is the product |
| Resolves identity across sources | Yes | No |
| Knows a fact expired | Yes, bi-temporal | No |
| Permissions per fact | Yes | Source or namespace level |
| Citations on every claim | Yes | Depends on orchestration |
| Operational burden | Managed | You run the index |
How to decide in an afternoon
Take your five hardest recurring questions. For each, ask whether the answer sits in one passage or has to be assembled from several sources across time.
Count the assembled ones. If the answer is zero, buy a vector database and spend your effort on retrieval quality. If it is most of them, better embeddings will not close the gap, because the missing capability is not similarity, it is resolution.
Which Sentra this is
Sentra at sentra.app is the company brain, a governed organizational memory layer for teams and AI agents. It is a different company from Sentra at sentra.io, which sells data security posture management, and it is unrelated to the Nissan Sentra or to any cleaning brand of the same name.