Trust Centre

Quality means no surprises.
Here is how we prove it.

The right memory, at the right time, no duplicates, no hidden data leakage, no surprises. An architecture you can inspect.

VEKTOR is self-funded, built by an independent developer with 30 years across computing, IT, data networking, storage, cybersecurity, and telecom. Memory for AI agents that centralize a person's most private conversational history locally with privacy by design. This page exists to show the alternative, and to make its guarantees proof you can check yourself rather than something we simply ask you to believe.

Four pillars, one architecture

Four properties of the same system, each one verifiable, not asserted.

01 · PRIVACY

Privacy-enhancing by design, not by policy

PET isn't a promise VEKTOR makes and asks you to trust; it's a property of the architecture itself. Embeddings are computed in-process, on your device, with a local ONNX model. There is no cloud copy of your memory.

  • Air-gapped by default: the design makes surveillance structurally impossible, not just contractually prohibited
  • Nothing to intercept: no memory API calls leave the machine to be logged, breached, or subpoenaed
"Built on a local-first, air-gapped design where surveillance is physically impossible, because we simply do not have access to your data."
02 · TRANSPARENCY

Transparent software, and transparent cost

Two kinds of transparency, and most companies only give you one. VEKTOR's architecture is published, not proprietary-black-box; its pricing is flat, not metered.

  • Software: the MAGMA graph, AUDN loop, and REM cycle are documented on /research, not asserted, described
  • Cost: one flat subscription, no add-ons, no upsell ladder, no usage-based billing surprise on next month's invoice
"No cloud, no add-ons, no upsell ladder: every core system VEKTOR ships with, bundled into one flat subscription."
03 · DATA

Ownership, control, migration, convenience

Data sovereignty usually costs you convenience: self-hosting, manual backups, DIY migration. VEKTOR's bet is that it shouldn't have to.

  • Ownership: your memory graph is a real SQLite file, on your machine, with a real file path
  • Control: never moved to mandatory cloud storage, never used to train models, price changes get 30 days' notice, export is always available
  • Migration: the file is portable to any machine, unaffected by our pricing or our continued existence; cancel any time and it stays exactly where it is
  • Convenience: one file to copy is your entire backup; no server to run
04 · QUALITY

Quality of the solution, and the software

AUDN (Add / Update / Delete / None) fires before every write, so contradiction handling is structural, not a cleanup job run later. Speed is part of that same quality bar, not a separate feature.

  • ~28ms average recall locally vs. 200–800ms typical for cloud vector-DB round trips
  • Why it's fast: in-process execution (no network round-trip), C-speed vector indexing, and HNSW search that avoids full-table scans as memory grows
  • Ongoing, not one-time: published benchmark methodology (LongMemEval), a public changelog, and a continuous REM cycle rather than an occasional maintenance pass

What do customers actually want from a memory layer?

"Will my agent still feel instant once its memory has months of history in it?"

Speed

Yes. ~28ms average recall locally, in-process execution with no network round-trip, staying flat as history grows.

"Who can access my data, where does it live, and can I get it out again?"

Data control & privacy

Only you. Your memory graph is a local SQLite file, air-gapped by design, with no cloud copy to access.

"Will my bill change if I use this more, or is the price the price?"

No hidden costs

The price is the price. One flat subscription, no add-ons, no upsell ladder, no usage-based billing.

"Can I get my data in and out without being locked into one vendor's format?"

Ingress, egress & embedding

Yes. Embeddings are computed locally and your SQLite file is a standard, portable format, not a proprietary lock-in.

"If I switch machines, providers, or tools, does my memory come with me?"

Ease of migration

Yes. Copy the file to any machine; cancel any time and it stays exactly where it is.

"Do I have to trust a claim, or can I actually verify it myself?"

Transparency

Verify it yourself. Architecture, benchmarks, and policies are published, not just asserted.

Six different questions. One answer underneath all of them: control over your own data.

What "quality" looks like in practice

Not an abstract claim: a side-by-side of the same conversation, memory-managed two different ways.

Naive recall
User: "What's my deploy process?"
Agent recalls 3 duplicate entries from 3 sessions ago
One entry is stale, describes a pipeline you replaced last month
Two entries contradict each other on the rollback step
Agent picks one at random, confidently
VEKTOR recall
User: "What's my deploy process?"
AUDN already resolved the contradiction at write-time
Old pipeline entry was superseded, not duplicated
One current, deduplicated fact returned
Agent answers with the version you actually use now

Proof, not hype

Numbers published on /research and /ecosystem, linked here, not restated, so there's one source of truth.

~28msavg local recall
200–800mstypical cloud vector DB
4-layerMAGMA graph
AUDNfires before every write
On the roadmap

We're working toward the most private, secure, and fastest agentic memory tooling available built on the best of open and closed software, so your data stays private and secure while staying technology-agnostic across LLM providers.

Verify it yourself

Every claim on this page links to the actual document behind it.

Our architecture is mathematically proven to protect your privacy.

VEKTOR Memory
See the architecture