Security and trust
Context has boundaries. Keep them.
A useful memory system needs to understand who is asking. Slashh applies permissions during retrieval, before any context is used to compose an answer.
The path of a request
The permission check happens first, not last.
Follow one question down the line. By the time a model reads a word, the request belongs to a person and the context has been trimmed to what they may see.
Identify the asker
Establish the person, or the scoped agent acting for them.
Check the context
Evaluate access before any retrieved content reaches the model.
Return the evidence
Compose the answer from permitted sources, citations intact.
Withhold the rest
Nothing the asker may not see reaches the model: not its title, not a snippet, not a count. The answer never reveals that it exists.
Controls you can inspect
Every request belongs to a person, or to an agent acting on behalf of one, and agent keys are stored hashed. That person’s access decides the context before anything is retrieved.
An agent sees its owner’s access, narrowed again by the scopes it was granted. Scope only ever narrows. Treat keys as secrets and grant only the access a workflow needs.
Every fact carries its source, when it was true and when it was learned. Facts are invalidated, not deleted, so a historical question can show what was known then.
Every call lands in an append-only, hash-chained log: who asked, on whose behalf, which tool, which documents, which access decisions. CSV export included. The preview may keep it in server memory, so check the retention limits.
Self-host it, or run it air-gapped. The open core is MIT-licensed and runs offline with no model API key. If you add a model provider, review the full source-to-model path before connecting sensitive systems.
Starting points for evaluating your sources, identity setup and deployment. Examples shown are illustrative.
See the boundary
One question. Context for the person asking.
The same question, asked by a team member and by a guest. Each answer is built only from what that person may see.
Asked by both“Where did we land on the project timeline?”
Team member
An answer supported by the shared overview and the team’s project conversation.
- Shared overview
- Team project conversation
Guest
An answer supported by the shared overview available to that guest.
- Shared overview
- Team project conversationWithheld before the model is called
The guest’s answer gives no hint that anything was withheld. Illustrative example.
Private preview
Clear about where we are.
Evaluate the current deployment against your requirements. A capability description is not a certification.
Is Slashh SOC 2 certified?
No. Evidence exports and control mapping can support a review, but they are not a SOC 2 certification or an independent attestation.
What should we review before a pilot?
Identity provisioning, source permissions, credential handling, model providers, hosting location, retention, deletion and audit-log durability. We will work through the configuration and gaps with you.
Can we assume production audit retention?
No. The preview access-log implementation may use server memory. Confirm durable storage and retention requirements before relying on it for compliance evidence.
Does the preview include SCIM provisioning?
SCIM provisioning is not a published preview commitment. Discuss identity lifecycle and offboarding requirements before starting a deployment.