Skip to main content

Delentia memory vs vector databases

They do different jobs. A vector database finds similar items fast. Delentia's agent memory exists for history and control. It does not claim to recall better.

CapabilityVector databaseDelentia memory
Fast similarity search at scale
Recall by meaning (opt-in, local embedding model)
Append-only, hash-chained history of every change
View the store as it was at an earlier point
Revoke a memory with a written reason
Erase one person's text by destroying its key
Detects an edited record
Multi-tenant isolation in a shared service
Hosted, managed service

This table is the author's reading of features, not a measurement. The measurement (against LangGraph and mem0 on one workload) is on the Benchmark page.

Is Delentia's memory a replacement for a vector database?

No. A vector database (Pinecone, Weaviate) stores embeddings and finds similar items quickly at scale. Delentia's agent memory is a small per-person store whose point is history and control: every change is an event in a hash-chained log, a memory can be revoked with a reason, the store can be viewed as it was earlier, and one person can be erased. If you need similarity search at scale, use a vector database.

Can it search by meaning?

Optionally. With a local embedding model switched on, Delentia found the right memory in the top three for 22 of 30 paraphrased questions, against 24 of 30 for mem0 and 7 of 30 for Delentia's default word matching. So it does not beat the alternatives at recall; its difference is the history and the erasure.

Does it make me compliant with the PDPA?

No. Software cannot make an organisation compliant. Delentia gives you tools: one person's memory text is sealed under a key that is destroyed on erasure (checked in the raw database file), and every change is an auditable event. Erasure does not yet cover the separate skills database or audit rows written before sealing was added.

What happened to the '8-dimensional schema' and the 74% compression?

They were design ideas and a formula, not measured results, and are withdrawn. Measured in bytes, storing changes saves 39% to 90% against full snapshots, and ordinary compression of snapshots does about as well.

The agent memory in detail