Capabilities

OmniGENIE (in development): interaction, memory and prepared actions under human approval

This page lists the capabilities of the GENIE technology on which OmniGENIE is based. Each row carries a status label. Available means the capability is active in the productive GENIE platform today; it does not mean an OmniOmni integration exists. See the legend on Purpose & Value.

Operational memory

Capability	Description	Status
Fact extraction	Typed facts (persona, episodic, instruction, value, open question, inferred belief) are extracted from conversation turns and linked to their source turn.	Available
Revision history	Facts are never overwritten in place. A correction creates a new fact and marks the previous one as superseded, so the information basis of a past action can be reconstructed.	Available
Sensitivity-scoped recall	Facts tagged as sensitive (for example health or business secrets) require a higher relevance bar before they are surfaced in a conversation.	Available
Memory graph	A visual graph of a user's own stored memory is generated asynchronously after each session. It is a derived view, not a source of truth.	Available
Correction and deletion of single facts	A specific memory can be corrected or deleted by voice or in the user interface, with a short grace period to undo.	Available
Full-account erasure	Erasure removes personal data across the complete data model in one transaction. An automated test enforces that every user-scoped table is covered. A hashed tombstone prevents accidental re-creation.	Available
Reflection pass	An optional background pass can surface higher-level patterns from stored facts.	Implemented, not enabled

Details, including provenance, sensitivity tiers, deletion and retention, are on Memory & Data Lifecycle.

Voice, telephony and text

Capability	Description	Status
Real-time voice interaction	Low-latency conversation over WebRTC with natural interruption (barge-in). If a user interrupts, only what was actually heard is retained in memory.	Available
Telephony	Inbound and outbound calls and SMS via a SIP trunk, with tenant-scoped assignment of phone numbers.	Available
Cascade and speech-to-speech pipelines	Speech-to-text, language model and text-to-speech run as a cascade by default; an alternative low-latency speech-to-speech mode is available.	Available
Automatic voice provider fallback	If the primary text-to-speech provider fails, the system falls back to a secondary provider for that turn. There is no equivalent automatic fallback for the language model.	Available
Structured call artefacts	Calls produce structured notes and follow-up tasks that persist after the call. Full meeting minutes or decision recaps are not produced today.	Available
Voice recognition of enrolled persons	Opt-in recognition of enrolled speakers with explicit consent capture and revocation.	Implemented, not enabled
Meeting companion	A real-time assistant for multi-speaker meetings. An architecture design exists; the capability is not built.	Roadmap

Tools, coordination and approvals

Capability	Description	Status
Tool invocation with sensitivity tiers	Every tool available to the assistant is tagged with a sensitivity level (low, medium, high) that gates how and when it can be invoked.	Available
Tool invocation logging	Tool calls are logged with a cryptographic digest of their arguments and an outcome status. Raw arguments are not stored in the log.	Available
Web research with source references	An optional research tool retrieves and cites external sources when a search provider is configured.	Implemented, not enabled
Graded agent routing	A routing design aligned to the Linux Foundation Agent-to-Agent specification returns one of five outcomes for a request: allow automatically, allow with summary, require approval, block, or escalate to a human.	Implemented, not enabled
Delegation model	A structured grant (grantor, scope, allowed and forbidden actions, approval mode, expiry, revocation), modelled on the OAuth token-exchange and rich-authorisation-request standards. Currently scoped to one person's assistant acting for another.	Implemented, not enabled
Single-use approval token	A short-lived, cryptographically bound approval token (inspired by OpenID CIBA) for specific change flows. It is not a general financial approval gate today.	Implemented, not enabled
Workspace integration	Connection to a user's Google Workspace account (mail drafts, calendar, files) with tokens designed to be stored encrypted.	Available
Group and guest modes	An optional group feature for shared artefacts without sharing private memory; a guest mode without persistent memory.	Implemented, not enabled

In development for OmniGENIE

The following capabilities are part of the OmniGENIE target design. They do not exist in code today.

Capability	Description	Status
Action tiers for financial operations	Observe, Draft, Recommend, Execute with approval; fully autonomous execution disabled for financial-market actions.	In development
Institution-level tenant isolation	Isolation at organisation level, in addition to today's per-user scoping.	In development
Enterprise identity federation	Sign-in through the institution's identity provider (OpenID Connect or SAML) instead of consumer OAuth.	In development
Approval through OmniPersona	Requesting and signing approvals for OmniGENIE-prepared actions through the existing Proof-of-Action flow.	In development
Legal hold	A retention override that blocks deletion where statutory retention obligations apply.	In development
Automatic circuit breaker	Automatic switch to read-only mode on anomalous error rates or tool-call patterns.	In development

From personal assistant to institutional layer

GENIE was built as a personal, German-language assistant. The table shows how its concepts translate into institutional terms, and which translations are backed by an existing capability.

Personal concept	Institutional reading	Backed by
Second memory	Policy-governed, purpose-bound operational memory	Fact extraction, revision history, sensitivity-scoped recall (Available)
"Knows you"	Authorised, purpose-bound context per account	Per-user data scoping (Available); institution-level isolation (In development)
Phone assistant	Disclosed voice interaction agent with rules, handover and records	Telephony, structured call artefacts (Available)
Personal loyalty	Operation aligned to the mandating principal under an explicit grant	Delegation model, graded routing (Implemented, not enabled)
Agent-to-agent	Controlled coordination between agents with identity, permission and reviewable handover	Graded routing (Implemented, not enabled); OmniOmni action tiers (In development)

No performance figures, accuracy rates, capacity limits or latency values are published on this page. Such values are set and measured per deployment together with the institution.