Values
What a memory is worth.
Every capability below exists because something breaks without it. This page states what each one does, why it matters, and — where a guarantee is conditional — exactly what the condition is.
Tier one
Sovereignty
Where your knowledge physically sits, and who can read it. The differentiator most evaluations end up turning on.
Data sovereignty Live
Documents, vector embeddings, the knowledge graph, uploaded files and usage metering live only on infrastructure you own. Chordian orchestrates and bills the box; it never holds the knowledge.
This turns three normally-procedural compliance objectives into structural ones. Residency becomes inherent — the data is in your region because the machine is. Data minimisation becomes true by design. And a legal request served on the vendor produces nothing, because there is nothing to produce.
The condition: this applies to own-server and own-cloud deployments. Our hosted tiers are a secure, isolated, content-aware SaaS — saying so is what makes the sovereignty claim credible.
Zero-content vendor Live
A total compromise of Chordian yields no customer memory, no SSH keys, no cloud credentials and no box encryption keys.
Install credentials are held only for the install window and purged on success — never written to the application database, logs or audit trail. The box’s WireGuard private key and its data-encryption key are generated on the box and never transmitted. There is nothing on our side to steal that unlocks your memory.
Defence in depth Live
Seven independent layers on a customer box, nine on the platform. Each is meaningful alone; together an attacker must defeat every one, and several sit on infrastructure the vendor does not control.
The layer worth noticing is mutual TLS. WireGuard encrypts the pipe; mutual TLS authenticates the caller. Without it, anything reaching the tunnel subnet could talk to the box unauthenticated. With it, a request arriving without a valid certificate is refused outright — and it is enforced from the box’s first boot, so there is no plaintext window during installation.
Air-gapped operation Validated
The box installs with zero outbound calls to public registries — the full image and platform bundle stream over the secure channel. Suitable for regulated and disconnected sites.
A sovereign box with no model egress configured fails loud: it refuses to process rather than quietly falling back to a public API. That failure mode is the difference between a sovereignty guarantee and a sovereignty claim.
The condition: air-gapped install is unconditional. Air-gapped processing requires bring-your-own or in-box model mode; in the default proxy mode, content transits our proxy to reach the model vendor.
Tier two
Knowledge quality
Most memory products accumulate whatever they are fed. Chordian also maintains what it holds — it notices when sources disagree, measures how good its own knowledge is, and compiles what it knows.
Entity coherence Live
The same person or company arriving from your CRM, inbox, chat and an enrichment provider becomes one entity, not four near-duplicates.
Duplicate entities are the silent killer of enterprise knowledge graphs: they fragment relationships, corrupt counts and make answers subtly wrong. Resolution runs continuously — on every write — using strong identifiers first and name similarity second, with an ambiguous middle band routed to a human rather than guessed.
Knowledge integrity Live
The industry default is last-write-wins. Your CRM says a deal is in negotiation; an email two weeks later says it closed; the later sync overwrites, and nobody is told the data ever disagreed. The discarded value vanishes and there is no decision trail.
Chordian records the disagreement instead — value, numeric and temporal conflicts detected inline during ingestion. A human reviews current against new, with each value’s source attached, and chooses. The choice is written back to the graph and audit-logged.
Detection is an alert, never a gate. Ingestion is never blocked or slowed — a sync does not fail because two sources disagree.
Knowledge health Live
A report card for the quality of the knowledge — freshness, coverage, connectivity and coherence, weighted into one A–F grade, scored nightly for the whole memory and for each department.
It answers a question nobody else answers: is the knowledge in here any good? A low freshness score usually means a connector stalled; low coverage means many entities and thin evidence; low connectivity means isolated entities. Three different fixes — and the grade alone would not have told you which.
Compiled knowledge Live
Asking about an account normally means searching hundreds of documents and re-deriving the answer every time — slow, an inference call per question, and subtly different on each asking.
A synthesis is a pre-built profile per entity, kept current in the background: overview, current relationship, key contacts, recent activity, risks, likely outcome. Read once, consistent every time — and always carrying its source count, because a summary is never ground truth.
Time-aware facts Live
Every fact carries two timestamps: when it was true in the world, and when Chordian learned it.
That pair answers a question an ordinary knowledge base cannot — what did we believe on 1 June, and when did we find out otherwise? For anything audited, the second timestamp is often the one that matters, because it establishes what you could reasonably have known at the time.
Tier three
Trust and control
Who may see what, provably — and the controls a security review will actually ask about.
Data perimeter Live
Secrets and personal data are detected, classified and masked before content is embedded or stored, and again before it reaches a model.
Masking is at the level of the individual value, not the surrounding chunk. The obvious implementation — drop any chunk containing a secret — destroys the knowledge along with the liability. Value-level masking keeps the sentence, the citation and the explanation, and removes only the credential.
If the scanner cannot run, ingestion fails rather than storing content unscanned. That is deliberate and not configurable.
Tenant isolation Live
Every read and write carries a tenant scope. It is not a convention — the helpers fail closed, the scope is always derived server-side from your credential and never read from a request body, and coverage is verified on every pull request by walking the syntax tree of each endpoint.
A read crossing a tenant boundary returns 404, not
403, so the existence of another organisation’s
data is never disclosed.
Cited answers Live
Four engines run in parallel — vector similarity, graph traversal, time-aware recall and an answer cache — and are reranked before a model writes anything.
On the streaming endpoint the source list is emitted before the first answer token. That ordering is a guarantee, not an implementation detail: a reader sees what the answer is grounded in before they read the prose, which is the difference between spotting a hallucination and discovering it later.
Architectural invariance Live
Twelve locked architecture decisions, each backed by a permanent test. A violation blocks the merge and the deployment.
This is the answer to “how do you know the isolation you are describing is still true in six months?” Most vendors answer with a policy document. Here the answer is a failing build.
What we do not claim
The list we keep because it is the useful one.
A security reviewer will find these anyway. Better they come from us.
No prompt-injection scanning
The perimeter prevents leaks; it does not detect adversarial prompts.
No content-blind storage
We read your content in order to index and reason over it. If the vendor must be unable to read it, the answer is an own-infrastructure deployment, not a cryptographic one.
No selective erasure yet
Whole-tenant purge exists; per-record deletion does not. It is the most commonly requested missing control and it is on the roadmap.
No SOC 2 attestation yet
The control mapping exists and the controls are real. The audit is an organisational programme that has not completed.
See which of these you actually need.
Most evaluations turn on two or three of the above. Tell us which, and we will show you exactly how they work.