Comparison
PersistMemory vs a vector database
A vector database stores and searches embeddings. Memory is what you have to build on top of one. This is an honest account of that gap, including the part where we use a vector database ourselves.
What a vector database is good at
Pinecone, Weaviate, Qdrant, Chroma and pgvector are excellent at what they do, which is storing vectors and returning the nearest ones quickly. There is nothing missing from them. We use pgvector underneath this product, so this is not a comparison between a database and a better database.
Where they actually differ
The gap is everything between a document and a searchable fact. Deciding what in a message is worth remembering, resolving that the person in the email is the person in the meeting, noticing that today's answer contradicts March's, keeping the authorisation that came with the source, and re-embedding when the model changes. A vector database does none of that, correctly, because none of it is its job. It is roughly the work of a small team for a year, and it is the reason this product exists.
| PersistMemory | a vector database | |
|---|---|---|
| Storing and searching vectors | Yes, on pgvector. | Yes, and that is the point of them. |
| Deciding what to remember | Extracted from the source into discrete memories. | You chunk and decide before inserting. |
| The same person, named three ways | Resolved to one entity across sources. | Three unrelated vectors. |
| Contradictions | Kept as conflicts, with both sides and their sources. | Two similar vectors, both returned, neither preferred. |
| Access rules | Carried from the source and enforced on recall. | Metadata filters you apply correctly every time. |
| Changing embedding model | Re-embedded in the background. | A migration you plan and run. |
Pick a vector database when
- You are building retrieval over a corpus you already control and understand.
- You want the storage layer and intend to own the logic above it.
- Your access model is uniform, so a filter is enough.
Pick this when
- You would rather not build entity resolution and conflict handling yourself.
- Content arrives continuously from places you do not control.
- You need recall to respect who was allowed to see the source.