Vector-Retrieval Memory
Memories embedded and stored in a vector index; recall retrieves by semantic similarity to the current task. The default architecture at volumes flat files cannot serve.
When to use
Large memory corpora (thousands of entries) where keyword and index lookup miss semantically-related content.
Specification
| Memory type | semantic |
|---|---|
| Storage | vector database or embedded index |
| Retrieval | semantic similarity search, task-triggered |
| Update method | agent-written or pipeline-extracted |
| Decay / staleness handling | Nothing native: similarity ignores age unless the schema carries dates and retrieval filters or weights by them. Undated vector memory is the ecosystem's most common staleness hazard. |
| Risk level / maturity | medium / established |
Privacy and safety
Embeddings leak content; treat the vector store with the same access control as the source text. Write-time redaction rules apply doubly.
Failure modes
- Similar-but-stale beats relevant-and-current without date weighting
- Retrieval quality degrades invisibly as the corpus grows
- Chunking decisions made early become load-bearing and unfixable
Assessment
Earns its complexity only at scale, and imports a hard obligation: similarity search has no concept of true, so freshness and verification must be added by the schema around it.
Evaluation method: Retrieval precision/recall against a labeled query set; staleness audit on retrieved results.