Compare

How Companion compares

Companion Core next to the other home AI computers people ask us about, and Companion Memory next to cloud assistants, agent memory SDKs and AI planners across 28 architectural dimensions. Every fact about another product comes from that company’s own site or documentation, and the sources are listed below.

At a glance

Five home AI computers, side by side

Prices are in US dollars as each company listed them on November 12, 2026. Companion Core comes in four sizes, from Core 64 at $2,992 to Core 192 at $9,108.

Companion Core compared with other home AI computers
AttributeCompanion CoreUmbrel HomeOlares OneGhost CoreLucebox Zero 495
Price$2,992 (Core 64) to $9,108 (Core 192); Core 128 is $5,428$599 for 1 TB, the cheapest model on sale$4,999$3,499$5,999 for 128 GB and 2 TB, up to $8,449 for 192 GB and 4 TB
Processor and graphicsAMD Ryzen AI Max+ 395 with Radeon 8060S graphics; Ryzen AI Max+ PRO 495 on Core 192Intel N150, 4 cores; no GPU listedIntel Core Ultra 9 275HX with an NVIDIA RTX 5090 Mobile GPUAMD Ryzen 5 9600X with an NVIDIA RTX PRO 4000 Blackwell GPUAMD Ryzen AI Max+ PRO 495 with an AMD Radeon AI PRO R9700 graphics card
Memory for AI models64 to 192 GB unified memory. Core 128: up to 96 GB as graphics memory. Core 192: about 160 GB for the GPU16 GB system memory24 GB on the GPU, plus 96 GB DDR524 GB on the GPU, plus 64 GB DDR5128 or 192 GB unified, plus 32 GB on the graphics card
Storage512 GB to 2 TB NVMe, by model1 TB, 2 TB or 4 TB NVMe2 TB NVMe1 TB NVMe2 TB or 4 TB NVMe
Local model sizes it states8B to 32B on Core 64 (70B-class at 4-bit); 120B-class on Core 128Not statedNot statedNames Qwen 3.8-Next, Qwen 3.8-27B, Gemma 4-31B and Muse-Glimmer-30BRuns DeepSeek V4 Flash (284B); publishes tokens per second for named models
Memory of your own dataCompanion Memory: the mail, notes, meetings and files you import, for agents you authorizeNot statedNot statedBuilds context over time from your apps, files, screen history and devicesNot stated
App store550+ one-click apps300+ apps200+ appsNo app store; a list of supported connectionsNo app store; one-click setup for 7 client apps
Free software for your own computerYes: Companion Hub for macOS, Windows and LinuxYes: umbrelOS, free for personal useYes: Olares OSNoNo; its inference engine is free and open source
Pool more than one boxYes: Hub PoolNot statedNot statedNot statedYes: cluster kit, $799 per machine
Warranty1 year parts, 2 years labor1 year2 years1 year1 year, parts and labor
FinancingAffirm, subject to credit approvalAffirm and Shop Pay installmentsAffirmNot statedNot stated
Ships toUS addresses onlyEvery country except sanctioned ones44 countriesNot statedCountries listed at checkout; shipping and duties in the price
Availability on November 12Delivered in about 10 days; Core 192 ships January 2027Ships nowIn stockFirst batch sold out; sign-up for the second batchPre-order paid in full; deliveries expected January 2027

Competitor specifications and prices are approximate and change often. Check the vendor’s own page before you compare. “Not stated” means the company’s product page does not say.

Sources, read November 12, 2026

Where Companion is different

Room for big models, and a memory that is yours

More memory for local models

Core 128 has 128 GB of unified memory, up to 96 GB of it usable as graphics memory, and Core 192 has 192 GB, about 160 GB of it reachable by the GPU. That is room for 120B-class models on Core 128.

Compare Companion models →

Companion Memory

Import the mail, notes, meetings and files you choose. Memory connects people, places, projects and decisions on your own disk, and agents get only the context you configure.

Explore Memory →

550+ one-click apps

Companion Hub installs apps and agents in one click, including local alternatives to the cloud tools you rent today. Hub is free on a Mac, Windows or Linux computer you already own.

Explore Hub →

Pool more than one box

Hub Pool lets apps on one Hub use models on another paired Hub, and sends each request to the least busy one, over Tailscale with no ports to open.

Read about Hub Pool →

Which one fits you

The honest short version

  • You want the largest local models and a private memory of your own records: Companion Core 128 or Core 192.
  • You mostly run models of about 30B or smaller and want the fastest answers: a dedicated 24 GB NVIDIA GPU, as in Olares One and Ghost Core, generally answers faster than unified memory. Core trades some of that speed for far more room.
  • You build with coding agents and want published speed figures: Lucebox Zero 495 adds a 32 GB graphics card to the same chip as Core 192 and publishes tokens per second for named models. It is a pre-order, with deliveries expected in January 2027.
  • You want a home server for files and apps on a budget: Umbrel Home, or Companion Hub free on a computer you already own, or Server on a stick for $120.
  • You live outside the United States: Umbrel, Olares and Lucebox ship internationally. Companion ships to US addresses only today.

Build it yourself

What Core adds to the same chip

Core 64, 64+ and 128 use the AMD Ryzen AI Max+ 395, the chip in a Framework Desktop you can build yourself, and NVIDIA’s DGX Spark is another 128 GB desktop. Companion Hub has free Linux installers for x86 and ARM computers, so you can try it on hardware like that yourself. What Core adds:

  • A tuned, assembled, burn-in-tested server that pairs wirelessly and sets up headless, like a printer.
  • Companion OS, Hub, Memory and local models set up to share one interface after one guided setup.
  • A 1-year parts warranty and 2 years of labor, and a support team you can email or call.

What leaves the box

Exactly what stays on a Core, and what uses the internet

Some home AI computers promise that no data ever leaves the house. Every connected device uses the internet for something, if only for updates, so here is the full list for a Companion Core.

Stays on your Core

  • Local model inference
  • Your files and app data
  • Companion Memory and the records you import
  • Optional on-device personalization, which does not contribute to a shared company model

Uses the internet

  • CI Account sign-in and device pairing
  • Software updates and the app catalog
  • Remote access, when you reach your apps away from home
  • Crash reporting, which is opt-out and on by default
  • Any outside service you connect an app or agent to

Personal AI and memory architecture

28-dimension personal AI and memory matrix

Most personal AI tools solve one slice of the problem: OS assistants summarize recent notifications on a phone, cloud agents run tasks in a remote browser, developer memory SDKs store chat snippets for custom bots, and calendar planners pack meetings into work hours. Companion Memory and Companion Hub combine six architectural layers on hardware you own.

Layer 1: Compute, residency and sovereignty (Dimensions 1 to 5)

Where personal records live, how models are hosted, and how compute scales across household devices.

Layer 1: Compute, residency and sovereignty (Dimensions 1 to 5)
DimensionCompanion Memory + HubApple IntelligenceMeta MusexAI Grok BotAgent Memory SDKs (Mem0, Zep, Letta)AI Planners (Motion, Reclaim, Sunsama)Local Appliances (Umbrel, Home Assistant, Open WebUI)
1. Data residency and ownership100% on user-owned hardware (PostgreSQL 18 + pgvector + PostGIS + TimescaleDB). Optional encrypted R2 backup; crash reporting is opt-out and on by default.On-device SQLite and Spotlight index, with stateless bursts to Apple Private Cloud Compute and optional ChatGPT handoff.Provider cloud storage across Meta infrastructure and cloud sandbox virtual machines.Provider cloud storage on xAI servers and X platform infrastructure.Managed SaaS cloud by default, or self-hosted database when deployed by a developer.Provider cloud SaaS databases synced with cloud calendar providers.100% local disk on a user-managed server or NAS.
2. Hardware envelope and local VRAM ceiling16 GB to 192 GB unified memory on Companion Core (up to 160 GB usable by the GPU on Core 192), or any Mac, Windows or Linux PC running Companion Hub.8 GB to 16 GB shared RAM on iPhone and iPad; up to 192 GB unified memory on Mac Studio.Cloud H100 and Blackwell clusters; 2 GB to 4 GB edge budget on Ray-Ban Meta glasses.Colossus cloud GPU cluster; thin client on mobile, web and Tesla vehicles.Library only; depends on the host server and external LLM API.Cloud SaaS only; browser and mobile clients.8 GB to 192 GB depending on the home server or desktop.
3. Multi-node LAN and tailnet compute poolingHub Pool over Tailscale routes inference requests to paired Hubs. Peer training offload code in Hub is experimental and not yet connected to Memory.Single-device execution with Private Cloud Compute fallback; no household Mac or iPhone compute pooling.Cloud-only scaling.Cloud-only scaling.Not included.Not applicable (cloud SaaS).Manual remote Ollama URL configuration; no automatic peer load balancing.
4. Offline and local-network operationLocal search, chat, HBPE 24-hour forecasting and planning run on your hardware; phone apps keep notes, tags, saved searches and compact search docs offline via PowerSync SQLite. Sign-in, updates, the app catalog and optional map/place lookups use the internet.Partial: ~3B on-device summarization and Siri intents work offline; complex reasoning requires Private Cloud Compute.Requires cloud connectivity for Muse reasoning and cloud VM workspaces.Requires cloud connectivity for Grok inference and live search.Works offline only if paired with a self-hosted database and a local LLM.Cached view only; scheduling solver requires cloud connectivity.Full local operation once models and containers are downloaded.
5. Authentication and session isolationTraefik forward-auth with HMAC-SHA256 timestamp verification (30s skew limit), one-time SpentHubAssertions replay ledger, and stable URN binding. Multi-user concurrent operation on one box is on the roadmap.Single Apple ID per device profile; Family Sharing shares selected albums and calendars rather than a shared AI server.Single Meta account per session.Single X or xAI account per session.Developer-supplied user_id or agent_id string filter.Workspace and team RBAC in cloud SaaS.App-by-app accounts or reverse-proxy logins.

Layer 2: Ingestion, normalization and whole-life memory graph (Dimensions 6 to 10)

How historical archives, messages, photos, locations and agent conversations are normalized into a queryable graph.

Layer 2: Ingestion, normalization and whole-life memory graph (Dimensions 6 to 10)
DimensionCompanion Memory + HubApple IntelligenceMeta MusexAI Grok BotAgent Memory SDKs (Mem0, Zep, Letta)AI Planners (Motion, Reclaim, Sunsama)Local Appliances (Umbrel, Home Assistant, Open WebUI)
6. Historical archive and live connector breadthAbout 80 registered archive importers (Google Takeout, Apple Health/Photos/Messages, Meta Facebook/Instagram/WhatsApp, Discord, Telegram, ChatGPT, Claude, Amazon, Spotify, Strava and more) plus 11 bulk /api/v1 capture lanes.Indexes Apple Mail, Messages, Photos, Calendar, Notes and Safari on the device; does not ingest third-party archives.Meta social graph, Messenger, Instagram, WhatsApp threads and Ray-Ban camera captures, plus files uploaded to a task.X posts, bookmarks and DMs, plus files uploaded in chat and optional Google Drive or calendar connectors.Ingests only chat turns and text documents pushed by a developer via API.Syncs active calendar events and task integrations (Google/Outlook Calendar, Slack, Linear, Notion).Siloed per app (Immich for photos, Paperless-ngx for PDFs, Home Assistant for sensors).
7. Canonical entity and relationship schema6 core record kinds (Person, Place, Thing, Content, Event, Space) backed by ci.entities single/class-table inheritance, typed ci.relationships in PostgreSQL 18, and 768-d HNSW ci_search.embeddings.On-device Spotlight Semantic Index and entity graph across Apple system apps.Workspace file tree and session memory items; no user-auditable personal database.Flat conversation history and vector chunks.Mem0 stores vector facts and optional Neo4j triples; Zep/Graphiti stores bi-temporal graph edges; Letta stores Core/Recall/Archival blocks.Tasks, projects, time blocks and calendar events only.Separate schemas in each self-hosted container; no shared cross-app entity graph.
8. Spatio-temporal stay-point and place resolutionPostGIS spatial indexing, GPS stay-point clustering, reverse geocoding, Home and Work suggestions after 3 weeks of routine, and next-place transition predictions (places_next).Significant Locations and Photos Places on-device; not exposed as an open MCP or SQL query API.Coarse location and camera context on glasses; no historical stay-point database.IP or mobile location context for local queries; no historical stay-point graph.No native PostGIS or stay-point clustering.Location strings on calendar events only.Home Assistant device tracker zones or Dawarich GPS logs in separate databases.
9. Deterministic and ML entity resolution44-stage ci.pipeline_state enrichment ledger with automated duplicate detection and merges for people, places and things, plus manual merge or dismissal review queues.On-device face clustering in Photos and contact card merge suggestions.Not applicable.Not applicable.LLM prompt-based node deduplication during write calls.Not applicable.Immich face clustering for photos only.
10. Governed agent memory (L0, L1, L2) and slot supersessionL0 session context + up to 24 lines of the next-24h HBPE forecast (/api/memory/context), L1 boosted keyword+vector + 1-hop graph recall (/api/memory/recall), L2 8-category writes with single-value slot supersession (lives_in, works_at, role_is, uses_primary, prefers), conflict detection (/api/memory/conflicts), path drift (/api/memory/drift) and grounded compaction (behind a setting).No external agent memory write or recall API.Session memory notes saved across tasks; no formal slot supersession or code-path drift API.Cross-session conversational memory toggle; no structured slot supersession or conflict ledger.Core focus of Mem0, Zep (bi-temporal invalidation) and Letta, without an integrated personal archive or HBPE forecast.Not applicable.Open WebUI flat user memory strings.

Layer 3: Inference routing, behavioral modeling and personal training (Dimensions 11 to 16)

How fast classification, 24-hour behavioral forecasting, walk-forward calibration gates, and background queues work together.

Layer 3: Inference routing, behavioral modeling and personal training (Dimensions 11 to 16)
DimensionCompanion Memory + HubApple IntelligenceMeta MusexAI Grok BotAgent Memory SDKs (Mem0, Zep, Letta)AI Planners (Motion, Reclaim, Sunsama)Local Appliances (Umbrel, Home Assistant, Open WebUI)
11. Fast classification and quick-decision routingEmbedding quick-decision API (/api/model-routing/decide and /rank) available for apps and agents, regex task activity classifier (taskL0Classifier), and opt-in HBPE keyword pre-filter (PR #3438 wires ingress pre-filtering into planning scans; trained System 1 models are planned).On-device ANE classifier and task-specific LoRA adapter router.Cloud routing between fast Muse models and multi-agent reasoning models.Cloud router between Grok 4 Fast and Grok 4 Heavy thinking modes.Every memory extraction and update call invokes an external LLM.Heuristic and solver-based scheduling rules.Single selected model per chat in Ollama or Open WebUI.
12. Behavioral time-series forecasting (24-hour rhythm)HBPE service (port 8002): forecasts the next 24 hours in 96 fifteen-minute slots across 8 activity classes (WORK, EXERCISE, SOCIAL, ERRAND, LEISURE, TRANSIT, SLEEP, UNKNOWN), plus weekly, due-item and seasonal forecasts.Siri proactive app and shortcut suggestions; no 96-slot 24-hour behavioral forecast or open forecast API.Reactive to user prompts and scheduled cloud reminders; no personal behavioral time-series model.Reactive to chat and X stream triggers; no personal behavioral time-series model.Stores historical facts; does not forecast future time-of-day activity distributions.Deterministic calendar packing within user-configured working hours; does not learn a probabilistic life rhythm.Home Assistant trigger-action automations; no learned behavioral transformer.
13. Per-user on-device model training and nightly adaptationPer-user HBPE hierarchical transformer pretrained as a masked autoencoder, adapted nightly with a rank-8 LoRA adapter (last 3 copies archived for manual restore), plus a per-day logistic regression head. Unsloth decision trainer code is an unwired scaffold.Apple trains global task adapters; users cannot train custom LoRA weights on their own history.User data may train Meta cloud models depending on region and settings; no user-owned adapter weights.X and chat history may train xAI global models unless opted out; no user-owned adapter weights.Prompt and graph engineering only; no per-user weight updates.No per-user neural weight training.Manual fine-tuning scripts only; no automated nightly LoRA lifecycle.
14. Walk-forward calibration and baseline serve gateD3 Serve Gate: serves the neural HMT checkpoint only when its Brier score beats the better of B1 (Yesterday) and B2 (Last-Week-Same-DOW) by 0.02 and ECE is within 0.02; the logistic head takes precedence when 200+ labeled slots fit.Internal Apple quality assurance before OS release; no user-visible per-person calibration scorecard.No user-visible forecast calibration gate.No user-visible forecast calibration gate.Offline benchmark suites (such as LoCoMo) only.No probabilistic forecast calibration.None.
15. Interactive chat priority and GPU resource governanceFour background LLM queues yield while a user chat reply runs; heavier enrichment stages start only inside a nightly window (01:00 for 300 minutes, server time).iOS and macOS kernel QoS preemption for foreground UI responsiveness.Managed by Meta cloud cluster autoscaling.Managed by xAI cloud rate limits and cluster scheduling.Depends on external API rate limits.Cloud SaaS request queues.Manual OLLAMA_NUM_PARALLEL settings; background jobs can contend with live chat.
16. Durable background job orchestration and enrichment ledgerpg-boss v12 with about 44 durable queues, a shared dead-letter queue, 62 recurring cron jobs, and a 44-stage ci.pipeline_state versioned per-record enrichment ledger.XPC and BackgroundTasks framework on-device.Ephemeral cloud container task runners per session.Cloud request workers and scheduled prompt triggers.Async webhook or background worker for memory ingestion only.Cloud calendar sync and rescheduling workers.Per-container cron jobs with no unified cross-app pipeline ledger.

Layer 4: Agentic actuation, MCP interop and security governance (Dimensions 17 to 20)

How external AI agents connect over MCP and REST, how capabilities are scoped, and how prompt-injection exfiltration is blocked.

Layer 4: Agentic actuation, MCP interop and security governance (Dimensions 17 to 20)
DimensionCompanion Memory + HubApple IntelligenceMeta MusexAI Grok BotAgent Memory SDKs (Mem0, Zep, Letta)AI Planners (Motion, Reclaim, Sunsama)Local Appliances (Umbrel, Home Assistant, Open WebUI)
17. Native semantic intent catalog and MCP server36 curated intents across 14 domains served over MCP (/api/mcp) and REST (/api/agent/intents) on main (38 with proposed-task accept/decline intents in open PR #3438), plus 112 GraphQL queries and 138 mutations for unscoped keys.App Intents across 13 system domains for Siri and Apple Shortcuts; not exposed as an external MCP server.Consumes external MCP tools inside Meta cloud workspaces; does not expose a local MCP server of your personal memory.Calls xAI server-side tools (X search, web search, code execution) and remote MCP tools in the cloud API.Exposes 3 to 6 CRUD tools (add, search, get, delete memories) over MCP.Proprietary REST APIs or narrow task-only MCP wrappers.Home Assistant exposes smart-home device controls over MCP.
18. Multi-dimensional API key scoping and capability enforcementOrthogonal area scopes (MCP, Intents, Memory, Capture, or unscoped full access) paired with capability tiers (read sees 26 tools, write sees 32 on main or 33 with acceptProposedTask, full sees all) enforced on every route.OS app entitlements and per-app privacy TCC prompts.OAuth user tokens stored in the cloud workspace.Account-level API keys and OAuth tokens.Project API key with user_id parameter.Workspace API keys or OAuth tokens.Per-app API tokens.
19. Intra-turn taint tracking and prompt-injection exfiltration gateTwo-layer defense: L0 context wraps recalled records in <<<UNTRUSTED_CONTENT fences, and PendingActionService tracks intra-turn followsLookup taint so any write or external tool call after a lookup queues as TOOL_QUEUED for human approval (per Autonomy setting; always queued for managed apps).On-device safety classifiers and user confirmation sheets before sending messages or mail.Sentinel L4-to-L7 network proxy inside the cloud VM blocks unauthorized domains and HTTP methods, paired with deterministic confirmation cards.Model-level safety alignment only; no intra-turn taint gate on personal memory reads.No prompt-injection fence or tool-execution taint gate.Not applicable.No intra-turn taint tracking.
20. Owner-only pending action approval and audit ledgerApprovals page (/approvals) queues sensitive tool calls (messaging_send, planning_completeTask, memory_forget, import_import_inbox), blocks managed apps from approving their own calls, and records approval decisions.Ephemeral Siri confirmation dialog with no persistent tool audit ledger.In-thread confirmation prompt in the cloud chat UI.No persistent owner approval queue.None.None.None.

Layer 5: Closed-loop daily planning, insights and surfaces (Dimensions 21 to 25)

How passive commitments become scheduled blocks, verified completions, grounded insights and daily reflections.

Layer 5: Closed-loop daily planning, insights and surfaces (Dimensions 21 to 25)
DimensionCompanion Memory + HubApple IntelligenceMeta MusexAI Grok BotAgent Memory SDKs (Mem0, Zep, Letta)AI Planners (Motion, Reclaim, Sunsama)Local Appliances (Umbrel, Home Assistant, Open WebUI)
21. Passive commitment extraction and trust-dial triageInboxScanService runs every 30 minutes over emails, texts and chat messages; TodoScanService runs daily at 06:00 over memories and user agent turns. Autonomy settings (cautious, balanced, hands_off) decide whether tasks wait in Proposed or open automatically.Found in Apps suggestions for events and reminders from Apple Mail and Messages.Extracts tasks only when prompted inside a chat thread.Extracts tasks only when prompted inside a chat thread.None.Manual task entry or explicit email-to-task forwarding.None.
22. Timezone-aware gap solver and calendar block placementPlannedBlockService runs a deterministic first-fit search using ci.users.timezone, workStart and workEnd (default 09:00 to 17:00), flexDays (default 7), and task estimate (default 30 min) to write planned_block events without overwriting timed events.Displays calendar events and reminders; does not automatically gap-solve unscheduled tasks into free calendar blocks.Can call calendar APIs when asked; no built-in interval gap solver.No built-in calendar gap solver.None.Core strength of Motion and Reclaim (cloud-hosted auto-scheduling solvers).Nextcloud Calendar or CalDAV storage without an AI gap solver.
23. Closed-loop completion verification from real-world evidenceHourly VerifyDone checks (hourly at :45) and 6-hour TodoProcessService scans cross-check active tasks against recent memories, messages and ended calendar blocks to suggest completion (or auto-complete on attributed evidence under hands_off), skipping previously declined suggestions.Manual checkbox completion in Reminders.Tracks only steps executed inside its own cloud browser or sandbox.None.None.Manual checkbox completion or time-elapsed block expiry; cannot read your texts, receipts or GPS to verify a task is done.None.
24. Computed metrics, anomalies and suggestion rules13 daily metrics are computed in SQL hourly with 28-day z-score, 90-day CUSUM and forecast-divergence anomaly detectors plus templated metric answers; separate LLM insight cards summarize graph changes and upcoming events, and suggestion rules adapt to accept/dismiss feedback.Journal suggestions from recent photos, workouts and locations.Ad-hoc data analysis when files are uploaded to a task.Real-time X trend summaries and news digests.None.Meeting-load and focus-time hour charts.Home Assistant energy and sensor graphs.
25. First-person reflections and 5-view spatial explorerReflection question catalog answered from your own history with confidence thresholds, plus a 5-view Explorer (Map, List, Timeline, Gallery, Graph) and review queues for people, places, things, tasks, asks and approvals.Apple Journal app and Photos Memories.Chat and document workspace UI.Chat and X feed UI.Developer JSON API only; no consumer UI.Calendar and Kanban board views only.Separate web UIs for each installed container.

Layer 6: Interpersonal consent, sharing and decentralized trust (Dimensions 26 to 28)

How questions from family, friends and peers are bounded by cryptographic identity and a disclosure ladder.

Layer 6: Interpersonal consent, sharing and decentralized trust (Dimensions 26 to 28)
DimensionCompanion Memory + HubApple IntelligenceMeta MusexAI Grok BotAgent Memory SDKs (Mem0, Zep, Letta)AI Planners (Motion, Reclaim, Sunsama)Local Appliances (Umbrel, Home Assistant, Open WebUI)
26. Decentralized identity and key verificationUses signed RFC 9421 asks with a did:keri:D key id convention or did:key today (when a public peer endpoint is set); full KERI (rotatable AIDs, Key Event Logs, pre-rotation, OOBI) and ACDC/SAID credentials are on the target design roadmap.Centralized Apple ID and iCloud Keychain.Centralized Meta account.Centralized X and xAI account.Developer-managed string identifiers.Google or Microsoft OAuth login.Local username and password or self-hosted OIDC.
27. Peer-to-peer coordination (DIDComm v2 and Local.Pub)RFC 9421 signed HTTP asks and Local.Pub profile lookup preview (behind a setting); DIDComm v2.1 messaging library blocks exist in the API and are not yet connected to a screen or socket.iMessage (PQ3 encryption) and AirDrop, locked to the Apple hardware ecosystem.Platform messaging on WhatsApp, Instagram and Messenger.X Direct Messages.None.Cloud booking links (such as Calendly-style pages).None built in.
28. Interpersonal ask-and-share disclosure ladderConservative-by-default Ask resolver across 4 peer trust tiers (Family, Trusted, Peer, Stranger) plus same-box Household, with a 6-rung disclosure ladder (none, boolean, count, summary, redacted, full), database grants, and an append-only ci.ask_audit ledger.All-or-nothing iCloud Shared Photo Library or Shared Calendar.Public or link-based sharing of a completed task output.Public or link-based sharing of a chat thread.None.Free/busy or full event details on shared calendars.File-level share links in Nextcloud or Immich.

Quick answers

Comparing home AI computers

Reviewed

How is Companion Core different from Umbrel Home?

Umbrel Home is a home server for apps and files: an Intel N150 with 16 GB of memory and no GPU listed, umbrelOS and 300+ apps, at $599 for the cheapest model on sale (Umbrel's site, November 12, 2026). Companion Core is built for local AI as well: 64 to 192 GB of unified memory for local models, Companion Memory for your own records, and 550+ one-click apps in Companion Hub. Core 64 costs $2,992. If you only need files and apps, Umbrel Home costs less, and so does the free Companion Hub on a computer you already own.

How does Companion Core compare with Olares One?

Olares One costs $4,999 and pairs an NVIDIA RTX 5090 Mobile GPU with 24 GB of memory and 96 GB of DDR5 (Olares's shop, November 12, 2026). Companion Core 128 costs $5,428 and has 128 GB of unified memory, up to 96 GB of it usable as graphics memory, so a model too large for a 24 GB GPU can still run entirely in graphics memory. Olares ships to 44 countries with a 2-year warranty. Companion ships to US addresses only, with a 1-year parts warranty and 2 years of labor.

How does Companion Core compare with Ghost Core?

Ghost Core costs $3,499 and pairs an NVIDIA RTX PRO 4000 GPU with 24 GB of memory and 64 GB of DDR5; its first batch was sold out on November 12, 2026 (Ghost's site). Companion Core 64 costs $2,992 with 64 GB of unified memory, and Core 128 costs $5,428 with 128 GB, room for 120B-class models. Both keep personal context on the device. Companion adds an app store with 550+ one-click apps, free Hub software for computers you already own, and Affirm financing.

How does Companion Core compare with Lucebox?

Lucebox Zero 495 is a headless inference workstation for developers: an AMD Ryzen AI Max+ PRO 495 with 128 or 192 GB of unified memory plus a 32 GB Radeon AI PRO R9700 graphics card, from $5,999 (Lucebox's site, November 12, 2026). It is sold as a pre-order paid in full, with deliveries expected in January 2027, and it publishes tokens-per-second figures for named models. Companion Core 128 costs $5,428 and is delivered in about 10 days. Companion adds Companion Memory, an app store with 550+ one-click apps, free Hub software for computers you already own, and Affirm financing.

Can I run Companion on hardware I already own?

Yes. Companion Hub is free for macOS, Windows and Linux, and Server on a stick ($120) carries Companion OS for an x86_64 PC with at least 16 GB of memory. Companion Core is the ready-to-launch option: assembled, burn-in tested, set up after one guided setup, and covered by a 1-year parts warranty and 2 years of labor.

What does a Companion Core send over the internet?

Local models, your files and Companion Memory run on the Core. CI Account sign-in, software updates and the app catalog use the internet, remote access does when you reach your apps away from home, and crash reporting is opt-out and on by default. Apps and agents also reach any outside service you connect them to.

How does Companion Memory compare with Apple Intelligence, Meta Muse and xAI Grok Bot?

Apple Intelligence runs a ~3B on-device model across Apple apps and sends larger requests to Apple Private Cloud Compute or an optional ChatGPT extension. Meta Muse and xAI Grok Bot run in provider clouds with cloud-hosted session state and tool sandboxes. Companion Memory runs on hardware you own (PostgreSQL 18 with pgvector, PostGIS and TimescaleDB), imports about 80 archive formats across Apple, Google, Microsoft, Meta, Discord, Telegram and WhatsApp, forecasts your next 24 hours locally in 96 fifteen-minute slots, and serves context to any MCP-compatible agent you authorize.

How does Companion Memory compare with developer memory SDKs (Mem0, Zep, Letta) and AI planners (Motion, Reclaim, Sunsama)?

Mem0, Zep (Graphiti) and Letta are developer SDKs for storing chat facts and temporal graph edges inside custom apps, without end-user archive importers, behavioral time-series forecasting, or a daily planner UI. Motion, Reclaim and Sunsama are cloud task and calendar planners without a personal knowledge graph, multimodal archive search, or an MCP server for local coding and chat agents. Companion Memory combines a local whole-life knowledge graph, an MCP and REST agent memory plane, and a daily planning and review loop on your own machine.