Ingest approved sources once
Register a file or text source, preserve its metadata, parse and chunk the content, and keep citation locations so answers can point back to evidence.
Back to products
Zelvora Brain
Governed organizational knowledge and memory.
Brain is a secure knowledge and memory service for human users and role-specific AI Employees. It centralizes what an organization knows, retrieves only authorized context, grounds answers in source evidence, and turns conversation into durable memory only through approval.
- 3
- Scope levels: org, employee, conversation
- 5
- Core backend services
- 2
- Transport contracts: REST and MCP
- 0
- Cross-tenant leakage target
Product vision
One knowledge layer, many digital employees.
Brain is a multi-tenant retrieval, knowledge, and approved-memory platform. A single organization can maintain shared knowledge, employee-specific knowledge, and short-lived conversation context without standing up a separate RAG stack for every digital employee.
The boundary is deliberate. Brain answers what this organization knows and which approved rule applies. It does not execute CRM writes, accounting actions, messages, document extraction, or report generation - so knowledge can be reviewed and secured independently from action-capable services.
Overview
Ingest, retrieve, ground, and govern.
Filter before you rank
Organization ownership, approval status, and employee or conversation scope constrain candidate retrieval before semantic ranking - never after.
Ground every answer
The answer layer receives only the authorized retrieval set and returns grounded text with citations, or states that evidence is insufficient.
Govern durable memory
Conversations may propose memory, but only approved memory becomes active production knowledge, with provenance and version history.
Backend services
A modular architecture behind one application layer.
Ingestion, retrieval, grounded answering, memory, and observability all share the same authenticated application layer regardless of transport.
Knowledge ingestion
Source registration, parsing, chunking, indexing, source versioning, and background jobs for large files processed in workers.
Scoped retrieval
Tenant- and role-constrained search, ranking, and citation metadata so retrieval security is part of answer correctness.
Grounded ask
LLM generation over retrieved evidence with explicit insufficient-evidence behavior instead of a forced confident answer.
Approved memory
Memory proposal, review, activation, rejection, and full audit history so chat never silently becomes production instruction.
Security and observability
Authentication context, tenant-safe cache, usage, latency, cost tracking, and audit events across every workflow.
Portable access layer
REST and MCP act as transport only. Both call the same application layer so Brain behaves identically regardless of client.
Data model
Ownership and traceability are explicit records.
Every retrievable object carries an organization identifier, a lifecycle state, and a scope. Chunks keep page or section locations so answers can point back to evidence.
Record
Purpose
Critical control
Source
Original document or text metadata and lifecycle
Organization ownership plus approved state
Chunk
Searchable content with source location
Same tenant and scope metadata as its source
Memory
Pending or approved durable knowledge
Approval status plus provenance
Job
Ingestion and indexing status
Tenant ownership plus resumable status
Cache
Reusable result metadata
Tenant, scope, and knowledge-version key
Security and reliability
Retrieval security is part of answer correctness.
Multi-tenant isolation, memory governance, and honest handling of insufficient evidence are treated as product controls rather than configuration details.
Candidate retrieval is already constrained by organization, approved status, and employee or conversation scope - Brain never searches a global collection and filters afterwards.
Cache keys include tenant, scope, and knowledge version. A key based on question text alone is unsafe because two organizations can ask the same question.
Cross-tenant tests use identical documents, identical queries, and similar employee names to prove sources, chunks, citations, memory, and cached answers never cross the boundary.
Retrieval quality and answer faithfulness are evaluated separately - good generation can hide poor retrieval, and strong retrieval can still be misrepresented.
Only approved memory becomes active, protecting the system from temporary instructions, user misunderstanding, and prompt-injection attempts.
Caller identity supplies the organization and allowed scopes. These values are never selected by the model.
Business use
Shared organizational context, not a single chatbot.
The same knowledge infrastructure supports many role-specific digital workers, while each employee receives only the knowledge and instructions permitted for that role.
Employee policy assistant
Staff ask policy, process, or product questions and receive grounded answers with source evidence.
Role-specific AI Employee
A finance, operations, support, or onboarding employee retrieves its approved SOPs and role rules before reasoning.
Customer-specific operations
An employee retrieves organization-specific rules such as date definitions, approval conditions, or reconciliation logic before another service executes.
Training and controlled learning
Teams train digital employees, propose reusable memory, and approve validated instructions for future work.
Cross-system workflow grounding
A runtime requests context from Brain, then invokes CRM, Kursiv, Reporting, or connectors separately.
Benchmark framework
Measure the whole system, not just the model.
Evaluation uses fixed organization-specific corpora with shared sources, employee-only sources, conversation-only sources, and deliberate distractors. The same question set runs across providers while corpus and scoring stay constant.
Retrieval quality
Precision and recall of the expected source chunks, plus ranking of authoritative evidence.
Answer faithfulness
Whether material claims are supported by retrieved sources and whether citations are correct.
Tenant isolation
Zero cross-organization retrieval, citation, cache, or memory leakage under adversarial tests.
Scope correctness
Employee-only and conversation-only knowledge stays inaccessible outside the permitted scope.
Latency and cost
p50 and p95 latency plus embedding, storage, retrieval, and model cost per grounded answer.
Memory integrity
Pending and rejected memory is never used; approved changes are versioned and auditable.
Quality, latency, and cost are reported together. Tenant isolation has one unconditional target - zero leakage. Other thresholds are established from real customer test sets before any performance claim is made.
Market comparison
A functional comparison, not a performance claim.
Enterprise AI knowledge is a mature market. Established vendors lead on search breadth, connector maturity, and deployment history. Brain competes on governance and scope.
Glean
Established strength. Enterprise search, permission-aware workplace knowledge, and a broad connector ecosystem.
Brain position. Compete less on generic search breadth; emphasize role-scoped digital employees and governed memory.
Microsoft Copilot Studio
Established strength. Agent building with Microsoft knowledge sources and deep enterprise ecosystem integration.
Brain position. Differentiate as a vendor-neutral Brain service with explicit scope hierarchy and portable contracts.
Salesforce Agentforce
Established strength. Agents grounded in Salesforce data and Data Cloud with native CRM action paths.
Brain position. Position Brain as system-agnostic organizational knowledge, with actions kept modular rather than CRM-native.
WRITER Knowledge Graph
Established strength. Enterprise generative AI with knowledge graph and RAG capabilities over trusted internal sources.
Brain position. Emphasize approval-based durable memory, action separation, and reuse across the product portfolio.
Differentiation worth proving
- Organization to employee to conversation scope hierarchy as a first-class concept.
- Approval and rejection lifecycle before temporary knowledge becomes durable production memory.
- Knowledge layer stays independent from action-capable CRM, ERP, and connector services.
- One Brain service reused across many AI Employee roles instead of a RAG stack per agent.
- Portable REST and MCP access that is not tied to a single SaaS ecosystem.
Primary users
Staff asking policy questions, plus finance, operations, support, and onboarding AI Employees
Core workflow
Register a source, index it under a scope, retrieve authorized context, answer with citations, and approve durable memory
Interfaces
REST and MCP transports over one shared application layer, portable across client ecosystems
Out of scope
CRM writes, accounting actions, messaging, document extraction, and report rendering stay in their own services
Differentiator
Approval-governed permanent memory combined with employee-level scope and knowledge/action separation