Skip to content

Where does N.E.K.O send conversations and memory?

N.E.K.O stores character memory locally by default, but relevant content can leave the device when you use a model provider, memory-processing task, free API route, cloud speech service, Steam Cloud, Workshop, online feed, browser feature, or remote Agent channel.

Last fact review: 2026-07-23.

Technical data-flow notice

This page explains the current implementation and official distribution statements. It is not a substitute for a jurisdiction-specific privacy policy, provider agreement, or legal review.

Data-flow overview

text
User input
├─ local conversation and per-character memory storage
├─ selected conversation or realtime provider
├─ optional memory-processing provider
├─ free API forwarding path
├─ optional telemetry
├─ optional proactive screen context
├─ user-triggered Steam Cloud or Workshop operation
└─ online content, browser, speech, or Agent service

What happens in each path

PathData involvedDestinationImportant boundary
Character memory storageRecent turns, facts, reflections, persona, journals, indexesLocal configured memory directoryLocal storage does not mean all processing is local
Conversation providerCurrent prompt, conversation context, attached inputSelected model providerProvider terms, retention, region, and account plan apply
Memory maintenanceRelevant conversation or memory textConfigured summary/extraction/correction providerRuns only for the applicable task, but can contain user content
Explicit memory recallSelected recalled snippetsActive conversation provider as tool outputThe full memory database is not automatically sent
Shipped free API pathInput needed for the free requestN.E.K.O forwarding service and service partnerCurrent Steam EULA distinguishes this from bring-your-own paid API use
TelemetryUsage and operational metadata described belowN.E.K.O telemetry serviceCan be disabled with environment variables
Proactive visionScreen stream or screenshots needed by the enabled featureLocal pipeline and configured vision/model pathPrivacy mode stops proactive viewing; manual screenshots are separate
Steam CloudAllowlisted character settings and memory filesSteam Auto CloudSnapshot is not a full memory-directory backup
Workshop publishUser-selected card, supported model files, preview, optional reference voiceSteam WorkshopPublication and asset licensing are user decisions
DEBUG diagnosticsQuery or tool arguments in some debug pathsLocal logs unless the user shares themDo not assume every log is content-free

Free API and your own provider key

The current Steam EULA describes two relevant paths:

  • with a paid provider API, input is sent from the device to the selected provider;
  • with the free API service, input can be forwarded through N.E.K.O servers to service partners.

The selected provider remains an independent data processor with its own terms. A provider entry existing in the configuration does not establish the same retention or privacy behavior across all providers.

Local memory and remote processing

The memory system keeps separate per-character layers for recent context, facts, reflections, and persona, backed by a chronological local database. Several maintenance tasks may use an LLM:

  • recent-history compression and review;
  • fact extraction and correction;
  • reflection synthesis and promotion;
  • persona merge and contradiction handling;
  • explicit recall returned to the active conversation model.

Optional embedding inference is local CPU ONNX, but that does not make the LLM-based maintenance tasks local. Read Memory System for the current runtime contract.

Telemetry

The repository README states that telemetry is enabled by default and collects operational categories such as:

  • model and call type;
  • token, request, and error counts;
  • application version, experiment information, locale, timezone, and distribution;
  • a pseudonymous device identifier and, in applicable Steam environments, a Steam numeric ID.

It states that raw conversation text, voice, images, API keys, email addresses, and phone numbers are not telemetry payloads. The implementation and the README must remain synchronized.

To opt out:

text
DO_NOT_TRACK=1

or:

text
NEKO_DO_NOT_TRACK=1

Screen and proactive-vision controls

Privacy mode stops proactive vision and releases its screen stream. It does not mean that a manual screenshot or a user-initiated screen-sharing action is technically impossible. First-run behavior can differ by distribution or region, so verify the current setting instead of assuming a universal default.

Agent and plugin capabilities have separate enablement and readiness controls. See Agent System and Task HUD System.

Steam Cloud is a partial character snapshot

Cloud Save uploads and downloads a character unit through Steam Auto Cloud. The allowlist includes common flat files such as recent memory, facts, persona, reflections, and time_indexed.db, but excludes current sharded archives, some metadata, recovery journals, and SQLite sidecars.

A download can replace local same-name character data and therefore uses confirmation, active-session handling, and a local operation backup. Read Cloud Save API before calling it a backup or migration solution.

Controls available today

ControlWhat it doesWhat it does not prove
Choose a providerChanges the service receiving the applicable requestThat every other feature uses the same provider
Disable telemetryStops the project telemetry pathThat third-party providers receive no requests
Enable privacy modeStops proactive screen viewingThat manual screenshots cannot be requested
Disable Agent channelsPrevents those channels from dispatchingThat chat or memory providers are local
Avoid Cloud / WorkshopAvoids those optional transfer pathsThat model APIs are offline
Delete a current characterRemoves current runtime character memory pathsThat every historical legacy directory or remote provider copy is removed

Comfortable with these data-flow boundaries? View N.E.K.O. on Steam. Review the selected AI provider's terms before sending personal content.