Skip to main content

TurboVec Is What Agent Memory Looks Like in Production

· 4 min read
John McMahon
John McMahon
CEO, Valkyr Labs Inc

Most agent memory demos stop at embeddings. They turn text into vectors, run a nearest-neighbor search, and call the result context.

That is a toy version of the problem.

Production memory has to answer harder questions. Who is allowed to see the record? Which business object did the evidence come from? Was the indexed source stale, unchanged, or newly refreshed? Can the agent cite a generated application object without inventing a custom evidence format? What should happen when the retrieval is weak, expensive, or unsafe?

That is the gap TurboVec is meant to close.

TurboVec is Valkyr Labs' product framing for GrayMatter's governed vector memory layer. Under the hood, the system is not just a loose vector store bolted onto chat. ValkyrAI already has a SemanticIndexEntry path, source-specific semantic reindexing, portable Java cosine scoring, generated-domain target search, and receipt-backed hybrid retrieval.

The important word is governed.

Vector Search Is Not Enough​

The common pattern is simple: chunk documents, embed chunks, search chunks. That works until the agent needs to operate inside real business software.

Business context is not just documents. It is MemoryEntry, ContentData, Task, Workflow, Product, Organization, Customer, Invoice, Application, and every other object the tenant can see through the live ThorAPI schema. If the memory layer flattens all of that into anonymous text, the agent loses the reason the data mattered in the first place.

GrayMatter's semantic index keeps target references attached. The code path can index generated target types, target IDs, tenant scope, searchable fields, relationship fields, taxonomy fields, source hashes, summary text, keywords, and vector bytes. That gives the agent a way to retrieve business objects as business objects, not just paragraphs.

Portable First, Accelerated Later​

TurboVec's practical advantage is portability. The baseline is not "you must buy a specialized vector database before memory works." ValkyrAI includes a portable path over SemanticIndexEntry using Java cosine scoring and float32 vector bytes. Optional accelerators can come later, but the correctness path works across ordinary ThorAPI deployments.

That matters for private AI automation. Teams want agent memory they can own, inspect, and run in their environment. A memory layer that only works when the data is exported to a separate hosted vector silo is a governance compromise before the first workflow starts.

Receipts Turn Search Into Evidence​

The strongest part is how the vector layer connects to Retrieval Receipts.

GrayMatter can route receipt-backed retrieval across sparse, dense/vector, graph, and summary providers, then preserve score details, provenance, policy, coverage, and answer gating. Generated-domain semantic targets can become receipt items with entity type and entity ID intact. That means an agent can cite a found object through the same audit path as a memory record.

This is the production shape:

  • Reindex source-specific business evidence.
  • Keep target type and target ID attached.
  • Search through ACL-aware semantic providers.
  • Fuse vector, keyword, graph, and summary evidence.
  • Emit a retrieval receipt.
  • Let the answer policy decide whether the agent can answer confidently.

That is very different from "grab the nearest chunks and hope."

The Buyer Pain​

AI platform teams are being asked to make agents useful without letting them become untraceable. They need RAG, but they also need access control, schema awareness, source freshness, tenant boundaries, cost controls, and audit trails.

TurboVec gives Valkyr Labs a sharper way to explain the vector piece of GrayMatter:

It is not a vector database pitch.

It is governed vector memory for agents that need to work inside real software.

The next step is to see TurboVec in the context of the full agent factory: GrayMatter for memory and receipts, ValkyrAI for workflow execution, ThorAPI for generated application surfaces, and ValorIDE for builder operations.