Competitive matrix & architectural strategy
Companion Memory and Companion Hub combine hardware ownership on a user-managed server or desktop with a relational and vector memory store in PostgreSQL 18, a per-person behavioral forecaster (hbpe-service), quick-decision and chat runtimes, scoped MCP tool governance with approval checks, and peer ask-and-share controls.
This page compares Companion Memory (main) and Companion Hub (dev) across 28 technical dimensions against four categories of software and appliances:
- Consumer OS and Cloud Personal Agents: Apple Intelligence (on-device ~3B models,
App Intents, and statelessPrivate Cloud Compute), Meta Muse (Muse Secure VMwith Sentinel L4-L7 network egress rules and markdown memory files), and xAI Grok Bot (cloud VM workspaces and scheduled tasks). - Developer Agent Memory SDKs:
Mem0,Zep/Graphiti(bi-temporal knowledge graph), andLetta(MemGPTcore/recall/archival memory blocks). - AI Auto-Schedulers and Daily Planners:
Motion,Reclaim.ai,Sunsama, andTodoist(Todoist MCP). - Local-First Appliances and Self-Hosted AI Tools:
Umbrel(umbrelOS),Start9(StartOS),Home Assistant,Open WebUI, andKhoj.
Every Companion Memory capability below uses the same status labels as Personas (Works today, Behind a setting, Not yet, and Planned), plus In review for changes in open Companion-Memory#3438 .
Architectural positioning
flowchart TB
subgraph Positioning["Personal AI & Memory Landscape"]
direction LR
subgraph CloudSilo["Vendor Cloud VMs & OS Silos"]
MUSE["Meta Muse<br/>Cloud Secure VM + Sentinel<br/>+ Markdown Memory"]
GROK["xAI Grok Bot<br/>Cloud VM + Scheduled Prompts<br/>+ X Connectors"]
APPLE["Apple Intelligence<br/>On-Device ~3B + PCC Enclave +<br/>App Intents (Apple Ecosystem)"]
end
subgraph DevFrameworks["Developer Memory Libraries"]
MEM0["Mem0<br/>Vector + Graph Fact CRUD"]
ZEP["Zep / Graphiti<br/>Bi-Temporal Knowledge Graph"]
LETTA["Letta / MemGPT<br/>Self-Editing Context Blocks"]
end
subgraph NarrowApps["Single-Job SaaS & Self-Hosted Silos"]
PLAN["Motion / Reclaim / Sunsama<br/>Calendar & Task SaaS"]
APPL["Umbrel / Start9 / Home Assistant /<br/>Open WebUI / Khoj"]
end
subgraph SovereignBox["Sovereign Personal Server + Memory Graph"]
CI["Companion Memory + Hub<br/>About 80 Importers · PostgreSQL 18<br/>HBPE 96-Slot Forecaster · 36 Intents<br/>Planning Loop · Peer Asks"]
end
endThree structural combinations on user-owned hardware
- Whole-Life Archive + Relational Memory Store (
PostgreSQL 18):- Companion Memory registers about 80 archive importers (
backend/apps/api/src/domains/import/datasources, Works today) alongside 11 bulk/api/v1capture lanes, document parsing viasummary-service:8002(MarkItDown), and a 44-stageci.pipeline_stateenrichment ledger (14promote,14index,11enrich,5clean). - Most cloud assistants and developer memory SDKs start with blank chat histories unless custom ETL connectors are written around them.
- Companion Memory registers about 80 archive importers (
- Per-Person 24-Hour Forecasting + Offline Planning Loop:
hbpe-service(:8002, Pythonci-hbpe) forecasts the next 24 hours in 96 fifteen-minute slots across 8 activity classes (WORK,EXERCISE,SOCIAL,ERRAND,LEISURE,TRANSIT,SLEEP,UNKNOWN) using a per-day logistic regression head or a per-person hierarchical transformer guarded by the D3 serve gate (Works today).- Scheduled NestJS jobs (
PlanningJobsCron->pg-bossplanningqueue) run inbox scanning every 30 minutes, morning to-do scanning at 06:00, completion verification every hour at:45, first-fit calendar block placement (PlannedBlockService), and SQL-first daily metrics across 13 scalars (Works today).
- Scoped Agent Interop (
MCP+/api/memory/*) + Approval Gate:- External tools connect over Streamable HTTP MCP (
/api/mcp, 36 curated intents across 14 domains onmain, 38 in open PR #3438 , plus250GraphQL tools for unscoped keys) and/api/memory/*(L0context,L1recall,L2governed writes with single-value slot supersession, conflict detection, path drift, and compaction Behind a setting). PendingActionServiceandtoolGate.tstrack intra-turnfollowsLookuptaint, wrap recalled blocks in<<<UNTRUSTED_CONTENTmarkers, queue sensitive calls on/approvals, and block installed apps from approving their own calls (Works today).
- External tools connect over Streamable HTTP MCP (
28-dimension multi-framework matrix
Each cell records the architectural mechanism, storage model, or shipped status on main (explicitly marking items that are Behind a setting, Not yet, Planned, or In review).
| # | Dimension | Companion Memory + Hub | Apple Intelligence (PCC) | Meta Muse (Secure VM) | xAI Grok Bot | Agent Memory SDKs (Mem0 / Zep / Letta) | AI Planners (Motion / Reclaim / Sunsama) | Local Appliances (Umbrel / Start9 / HA / Open WebUI) |
|---|---|---|---|---|---|---|---|---|
| I. | Compute, Residency & Sovereignty | |||||||
| 1 | Data residency and ownership | User-owned hardware (PostgreSQL 18 with pgvector, TimescaleDB, PostGIS, object store, Ollama). Crash reporting is opt-out (on by default); place lookups, map tiles, and first-time Hugging Face fallback model download use the internet unless overridden. | On-device index with stateless bursts to Apple Private Cloud Compute; optional ChatGPT handoff. | Meta cloud infrastructure and persistent cloud VM disks. | xAI cloud servers and X platform storage. | Managed cloud SaaS by default, or developer self-hosted DB. | Vendor multi-tenant cloud SaaS synced with calendar providers. | User-owned server, NAS, or desktop disk. |
| 2 | Multi-node local compute pooling | Hub Pool over Tailscale routes inference requests across paired Hubs (Works today). Hub peer training offload code exists in Hub (Not yet; not wired to Memory, image lacks ci-hbpe). | Single-device execution with PCC fallback; no household LAN pooling. | Cloud-only scaling. | Cloud-only scaling. | Not included. | Not applicable (cloud SaaS). | Manual remote OLLAMA_HOST configuration; no automatic pool load balancing. |
| 3 | Offline & local-network operation | Search, chat, HBPE forecasting, and planning run locally (Works today). Phone apps keep notes, tags, saved searches, and mobile_search_documents offline in SQLite via PowerSync; full records and embeddings stay on the box. | Partial: ~3B on-device summarization and Siri intents run offline; larger queries need PCC. | Requires cloud connectivity. | Requires cloud connectivity. | Offline only with self-hosted DB + local LLM. | Cached UI only; solver needs cloud connectivity. | Local operation once models and containers are pulled. |
| 4 | Data portability and export | Account export/import archive APIs (domains/portability) + direct PostgreSQL tables and object store files. | Per-app iCloud exports; semantic Spotlight index is non-exportable. | Editable markdown files + workspace download. | Chat history export; no relational memory graph export. | Developer-controlled database and JSON exports. | CSV task and .ics calendar exports. | Raw Docker volume mounts per self-hosted app. |
| 5 | Authentication & session isolation | Traefik forward-auth to Hub with HMAC-SHA256 timestamp check (<= 30s), SpentHubAssertions replay ledger, and stable urn:ci-hub binding (Works today). Multi-user concurrent household operation on one box is roadmap. | Single Apple ID per device profile; Family Sharing shares media and calendars. | Single Meta account per session. | Single X / xAI account per session. | Developer-supplied user_id / agent_id string key. | Cloud workspace RBAC. | Per-container accounts or reverse-proxy logins. |
| II. | Ingestion, Normalization & Whole-Life Memory Graph | |||||||
| 6 | Historical archive importers | About 80 registered importers (datasources/index.ts, Works today), including Google Takeout (Gmail, Photos, Location, Keep, YouTube, Drive/Purchases), Apple (Health, Photos, Music, Contacts, iMessage, BlueBubbles, macOS Activity), Meta (Facebook, Instagram, WhatsApp), Discord, Telegram, Twitter/X, LinkedIn, Reddit, ChatGPT, Claude, Gemini, Grok, Notion, Evernote, Kindle, Amazon, Spotify, Strava, Garmin, Fitbit, Oura, and Withings. | Indexes Apple Mail, Messages, Photos, Calendar, Notes, and Safari on-device; does not import third-party archives. | Starts from blank memory files plus manual uploads and connected OAuth channels. | Starts from blank history plus manual uploads and X / cloud connectors. | None built in; requires developer ETL code. | Syncs active calendar and task tools only. | Siloed uploads per app (Immich photos, Paperless-ngx PDFs, Khoj docs). |
| 7 | Live capture and ingestion lanes | 11 bulk /api/v1 lanes (events, communications, contents, things, inventory, browser-activity, notes, positions, receipts, product-models, observations) + POST /api/v1/files, web UI, mobile apps, and browser/desktop capture (Works today). | Continuous indexing of Apple system apps. | Cloud OAuth + WhatsApp / Instagram / Messenger / Telegram / Discord / Slack bridges. | X stream + Gmail / Calendar / Drive / messaging connectors. | Chat/text write API only. | Google/Outlook Calendar, Slack, Jira, Linear, Notion sync. | Home Assistant sensors + container-specific uploads. |
| 8 | Canonical entity and graph schema | STI/CTI root at ci.entities (Person, Place, Thing/Item, Content, Event, Space/Concept, File) + typed ci.relationships in PostgreSQL 18 + ci_search.vectors (GIN) + ci_search.embeddings (768-d nomic-embed-text HNSW) (Works today). | On-device Spotlight semantic index and per-app App Entity schemas. | 4 flat markdown files (MEMORY.md, USER.md, SOUL.md, HEARTBEAT.md). | Flat editable fact list + vector chunks. | Mem0 vectors + optional Neo4j triplets; Zep/Graphiti bi-temporal edges; Letta core/recall/archival text blocks. | Tasks, projects, and calendar time blocks. | Isolated per-container databases and vector collections. |
| 9 | Document parsing & chunk provenance | MarkItDown via summary-service:8002 /parse -> ci.file_chunks via DocumentChunkerService (Works today). | On-device Spotlight file parsing. | Cloud sandbox file reader (PDF/DOCX/XLSX). | Cloud sandbox file reader. | Requires external loader (LangChain / LlamaIndex). | File attachments only. | Built-in RAG chunkers (Open WebUI, AnythingLLM, Khoj). |
| 10 | Entity resolution & governed agent memory | clean.duplicate_detect -> ci.merge_candidates with auto-merge rules + manual merge/dismissal for people, places, and things (Works today). /api/memory/write supports 8 categories, 5 single-value supersession slots (lives_in, works_at, role_is, uses_primary, prefers), negation conflict flags (/conflicts), and /drift (Works today); compaction into KG is Behind a setting. | Contacts/Photos merge suggestions on Apple devices; no external /api/memory conflict or slot API. | LLM edits markdown lines in place; no conflict or path-drift API. | Flat memory edits in settings; no slot supersession ledger. | Zep/Graphiti bi-temporal edge invalidation (t_valid/t_invalid); Mem0 vector UPDATE/DELETE; no integrated personal archives or merge review queues. | None. | Open WebUI flat user memory notes only. |
| III. | Inference Routing, Behavioral Modeling & Personal Training | |||||||
| 11 | Fast scoring & System 1 routing | Quick-decision API (/api/model-routing/decide, /rank, embedding cosine similarity + trigram fallback) Works today for external apps/agents (unused internally on main); regex taskL0Classifier Works today; HBPE keyword fast path (HBPE_ENABLE_DECISION_MODELS=false) is Behind a setting; HBPE /decision/* routes are Not yet on main (In review in PR #3438 for planning ingress pre-filtering). | On-device Apple Neural Engine classifiers and task-adapter routing. | Cloud routing across Muse models. | Cloud routing across Grok Fast / Heavy modes. | Every extraction/update triggers an external LLM call. | Heuristic scheduling constraints and MILP/greedy solvers. | Single selected model per chat in Ollama or Open WebUI. |
| 12 | Behavioral time-series forecasting | hbpe-service:8002 forecasts the next 24h in 96 fifteen-minute slots across 8 L0 classes (WORK, EXERCISE, SOCIAL, ERRAND, LEISURE, TRANSIT, SLEEP, UNKNOWN), plus week-ahead recurrence (/forecast), due/overdue items (/forecast/due), and seasonal (/forecast/seasonal) (Works today). | Siri proactive app/shortcut suggestions; no 96-slot forecast horizon or API. | None (scheduled prompt crons only). | None (scheduled prompt crons only). | Stores past facts; does not forecast future time-slot distributions. | Deterministic task packing within work hours; no learned behavioral rhythm model. | Home Assistant trigger automations; no learned daily transformer. |
| 13 | Per-user training & nightly adaptation | Per-day multinomial logistic regression head (40 features) + per-person hierarchical transformer (256d, 6L, 8H) pretrained with MAE and adapted nightly (02:30) via rank-8 LoRA on yesterday’s tokens, keeping 3 archived copies (lora_<uuid>.pt.vNNNN) for manual restore (Works today). Note: UnslothDecisionTrainer and Hub TrainingModule are Not yet (unwired scaffolds); Qwen3.5-0.8B (Q8_0 GGUF) runs as the local fallback activity labeler inside hbpe-service. | Global adapters shipped with iOS/macOS; no per-user weight training on appliance. | No user-owned adapter weights. | No user-owned adapter weights. | Prompt/graph storage only; no weight training. | No neural weight training. | Manual fine-tuning scripts only. |
| 14 | Calibration & D3 serve gate | D3 Serve Gate (serve.py): serves neural HMT only when its Brier score beats the best of B1 (Yesterday) and B2 (Last-Week-Same-DOW) by 0.02 and ECE is within 0.02; logistic head takes precedence when 200+ labeled slots exist (Works today). | Internal Apple QA prior to OS release; no per-user calibration gate. | None. | None. | Offline benchmarks (such as LoCoMo) only. | None. | None. |
| 15 | Interactive chat priority & GPU governance | Chat preemption: 4 background Ollama queues yield while chat.generate (llm queue) runs; non-index/embed enrichment stages start only inside the 01:00-06:00 server-time window (300 minutes, 90 min dispatch) (Works today). | OS kernel QoS prioritizes foreground UI. | Managed in Meta cloud clusters. | Managed in xAI cloud clusters. | External API rate limits. | Cloud SaaS queues. | Manual OLLAMA_NUM_PARALLEL tuning; background jobs can compete with chat. |
| 16 | Durable queues & enrichment ledger | pg-boss v12 (~44 queues + shared dead-letter queue), 62 recurring @Cron jobs (17 worker-only), and 44-stage ci.pipeline_state versioned ledger (SmartRotation scheduler is Behind a setting) (Works today). | iOS/macOS XPC and BackgroundTasks. | Ephemeral per-session cloud task workers. | Cloud request and cron workers. | Simple ingestion worker queue. | Cloud calendar sync workers. | Per-container cron jobs with no shared ledger. |
| IV. | Agentic Actuation, MCP Interop & Security Governance | |||||||
| 17 | Semantic intent catalog & MCP surface | 36 curated intents across 14 domains served over /api/mcp and /api/agent/intents on main (38 in open PR #3438 adding planning_acceptProposedTask and planning_declineProposedTask), plus 250 GraphQL tools (112 queries + 138 mutations) for unscoped keys (Works today). | App Intents across 13 OS domains for Siri and Shortcuts; not an open external MCP server. | Cloud VM consumes external MCP tools; does not expose your memory as a local MCP server. | Server-side xAI tools + cloud connectors. | Exposes 3-6 memory CRUD tools over MCP. | Proprietary REST APIs or task-only MCP wrappers. | Home Assistant exposes smart-home controls over MCP. |
| 18 | Orthogonal API scopes & capabilities | Area scopes (MCP, Intents, Memory, Capture, or unscoped full access) + capability tiers (read sees 26 curated tools, write sees 32 on main / 33 in #3438, full sees all) enforced on every route (Works today). | OS app entitlements + per-app TCC privacy prompts. | Cloud OAuth tokens in VM. | Account-level API keys / OAuth. | Project API key + user_id parameter. | Workspace API keys / OAuth. | Per-container API tokens. |
| 19 | Prompt-injection taint tracking & fencing | /api/memory/context and chat wrap retrieved records in <<<UNTRUSTED_CONTENT fences; PendingActionService and toolGate.ts track intra-turn followsLookup taint so subsequent writes or third-party calls are queued per Autonomy setting (or always for managed-app) (Works today). | On-device safety classifiers + user confirmation sheets. | Sentinel L4-L7 eBPF proxy blocks unauthorized egress from the cloud VM + confirmation cards. | Model safety alignment only; live browser takeover relies on user vigilance. | Raw recalled strings placed directly into LLM prompt. | Not applicable. | None built in. |
| 20 | Approval cards & self-approval block | /approvals (ApprovalsPage.tsx) holds TOOL_QUEUED calls (messaging_send, planning_completeTask, memory_forget, import_import_inbox, and tainted writes) and blocks managed-app keys from approving their own calls (Works today). | Ephemeral Siri sheet; no cross-agent approval queue UI. | In-thread confirmation card + Sentinel activity log. | Live screen-takeover prompt for browser logins/CAPTCHAs. | None. | Minimal draft confirmation in Stuff / Sunsama. | None. |
| V. | Closed-Loop Daily Planning, Insights & Surfaces | |||||||
| 21 | Commitment extraction & trust dial | InboxScanService (every 30 min) scans emails/SMS/chat (excluding agent turns) and TodoScanService (daily at 06:00) scans memories and user agent turns; Autonomy (cautious, balanced, hands_off) governs whether tasks wait in Proposed (balanced opens at >= 0.55) (Works today). | Siri “Found in Apps” suggestions for Mail and Messages. | Manual prompt or cron instruction only. | Manual prompt or cron instruction only. | None. | Email forwarding or manual sidebar pull (Sunsama). | None. |
| 22 | Calendar gap solver & block placement | PlannedBlockService first-fit gap solver uses ci.users.timezone, workStart/workEnd (09:00-17:00 default), flexDays (7 default), and defaultEstimatedMinutes (30 default) to place planned_block events without overwriting timed events (Works today). | Displays Reminders/Calendar side by side; no automatic task gap-placement. | Can call calendar APIs in chat; no built-in constraint gap solver. | None. | None. | Core strength of Motion & Reclaim.ai (multi-task auto-scheduling solvers). | Passive CalDAV storage (Nextcloud). |
| 23 | Real-world completion verification | VerifyDone (hourly at :45) and TodoProcessService (every 6h) inspect recent memories, messages, and ended blocks (DONE_PHRASES / CANCEL_PHRASES), append notes, propose completion (hands_off completes automatically on attributed evidence at >= 0.8), and respect declined hints (Works today). | Manual completion checkboxes. | Tracks only steps run inside its own cloud sandbox. | None. | None. | Manual checkbox completion or time-elapsed expiry; cannot read messages or receipts to verify completion. | None. |
| 24 | Computed metrics, anomalies & insights | 13 daily metrics computed in SQL hourly at :10 with 28-day z-score, 90-day CUSUM, and forecast-divergence anomaly checks plus TemplateNarrator metric answers; separate LLM insight cards summarize graph changes and events (3h); suggestion rules rank via empirical Beta arm mean (Works today). | Apple Health trends (single domain) + Journal suggestions. | Ad-hoc chat analysis on uploaded files. | X trend summaries and news digests. | None. | Weekly focus/meeting-hour charts. | Home Assistant energy/sensor charts. |
| 25 | Explorer views, reflections & review queues | 5-view Explorer (Map, List, Timeline, Gallery, Graph), confidence-gated reflection catalog (/self/reflections), Brief (/today, Behind a setting), Forecast (/forecast, Behind a setting), and review queues (Tasks, People, Places, Things, Asks, Approvals) (Works today). | Apple Journal + Photos Map/Memories; iOS/macOS widgets. | Chat + workspace file UI; messaging channels. | Chat + X UI; messaging channels. | Developer JSON API only; no consumer UI. | Calendar + task boards + iOS/Android/menu-bar widgets. | Separate container web UIs + Home Assistant companion widgets. |
| VI. | Interpersonal Consent, Sharing & Decentralized Trust | |||||||
| 26 | Cryptographic peer identity & KERI roadmap | Signed RFC 9421 HTTP asks accept Companion Memory’s did:keri:D<ed25519_x> key id convention (or Portal tokens) when CI_PEER_MEMORY_ENDPOINT is configured (Works today; non-CESR). Full KERI (E... AIDs, Key Event Logs, pre-rotation, OOBI), ACDC grants, and SAID digests are Planned (CI-Hub froze its KERI code Oct 30, 2026). | Centralized Apple ID & iCloud Keychain. | Centralized Meta account. | Centralized X / xAI account. | Developer string IDs. | Google / Microsoft OAuth login. | Local credentials or self-hosted OIDC. |
| 27 | Peer messaging (DIDComm) & Local.Pub | Local.Pub public profile lookup + did:key peer linking preview (Behind a setting: CI_PEER_LOCAL_PUB_ENABLED). DIDComm v2.1 encryption/packaging is built as tested API library code (Not yet wired to routes/sockets). | iMessage (PQ3) and AirDrop (Apple devices only). | WhatsApp / Instagram / Messenger. | X Direct Messages. | None. | Booking links (Calendly-style). | None built in. |
| 28 | Ask-and-share disclosure ladder & audit | Conservative Ask resolver across 4 peer tiers (stranger -> boolean, peer -> summary, trusted -> redacted, family -> full) plus household (same box only), 6 disclosure rungs (none, boolean, count, summary, redacted, full), min(stamped, current) monotonic revocation, folded asker status (answered / delivered), DB grants, ci.ask_audit, and Saved Searches (Works today). | All-or-nothing iCloud Shared Library / Calendar. | Shareable link to task output. | Shareable link to chat thread. | None. | Free/busy vs full event calendar shares. | Container-level share links. |
Deep-dive comparisons
1. Compute, residency and execution plane: Physical Hub vs. Cloud VMs vs. PCC
| Mechanism | Apple Intelligence (Private Cloud Compute) | Meta Muse (Secure VM) & xAI Grok Bot | Companion Memory + Hub (v2026.11) |
|---|---|---|---|
| Where background work runs | Apple Silicon device OS (BackgroundTasks) for indexed mail/messages; PCC servers are stateless request enclaves | Always-on Linux VM in Meta or xAI cloud hosting bash, browser GUI, and scheduled prompts | User-owned server or desktop running Companion Hub + pg-boss v12 (~44 queues, 62 @Cron jobs) + PowerSync SQLite subset offline on mobile |
| What requires an internet connection | PCC escalations, ChatGPT extension, iCloud sync | All cloud reasoning, VM automation, and OAuth connectors | CI Account sign-in, software updates, app catalog, default Portal place lookups, map tiles, first-time Hugging Face fallback download, and connected external services |
| Multi-machine scaling | Single device + Apple PCC only | Cloud GPU autoscaling | Hub Pool over Tailscale routes inference across paired Hubs (Works today); Hub peer training offload code is Not yet wired to Memory |
2. Memory, belief governance and quantitative self-knowledge
| Capability | Mem0 | Zep / Graphiti | Letta (MemGPT) | Meta Muse (MEMORY.md) | Companion Memory |
|---|---|---|---|---|---|
| Primary representation | Extracted fact vectors + optional Neo4j triplets | Bi-temporal graph (Episodic, Entity, Community) | Self-edited Core, Recall, and Archival blocks | 4 markdown files (MEMORY.md, USER.md, SOUL.md, HEARTBEAT.md) | Relational graph (ci.entities + ci.relationships), ci_search.embeddings/vectors, governed ci.agent_memories + ci_derived daily metrics (Works today) |
| Contradictions & code drift | Vector overwrite on UPDATE | Bi-temporal t_valid/t_invalid edges | Text block edits | In-place markdown edits | Single-value slot supersession (lives_in, works_at, role_is, uses_primary, prefers), negation conflict flags (GET /api/memory/conflicts), and file-path drift (POST /api/memory/drift) (Works today) |
| Whole-life archives | None (chat/text API) | None (episodes API) | None (chat/docs API) | Manual uploads + OAuth channels | ~80 registered importers + 11 /api/v1 capture lanes + PostGIS stay-point places (Works today) |
| Quantitative metrics | None | None | None | Cannot compute SQL regressions | 13 daily SQL metrics computed hourly with 28d z-score / 90d CUSUM anomalies + templated metric answers (Works today) |
3. Agent actuation, MCP governance and anti-exfiltration controls
| Security & Actuation Mechanism | Apple Intelligence (App Intents) | Meta Muse (Sentinel + Secure VM) | Companion Memory (domains/tools + domains/mcp) |
|---|---|---|---|
| Intent & tool surface | 13 OS schema domains (App Intents for Siri and Shortcuts) | Untyped MCP tools + bash + Chromium GUI clicks | 36 curated intents on main (38 in PR #3438 ) across 14 domains over /api/mcp & /api/agent/intents, plus 250 GraphQL tools (112 queries + 138 mutations) for unscoped keys |
| Capability & scope enforcement | OS app entitlements + TCC prompts | Sentinel L4-L7 eBPF proxy blocks unauthorized egress domains/methods | Orthogonal key scopes (MCP, Intents, Memory, Capture, or unscoped) + capability levels (read 26, write 32 on main / 33 in #3438, full all) enforced across MCP, REST, and internal tool calls |
| Prompt-injection & lookup taint | Safety classifiers + confirmation sheets | Sentinel egress rules + isolated cloud VM | <<<UNTRUSTED_CONTENT fences on L0/chat context + intra-turn followsLookup taint (toolGate.ts & PendingActionService) holding writes and external calls per Autonomy setting (always for managed-app) |
| Approval workflow & self-approval block | Ephemeral Siri sheet | Deterministic in-thread approval card | /approvals (ApprovalsPage.tsx) holds TOOL_QUEUED actions (messaging_send, planning_completeTask, memory_forget, import_import_inbox, tainted writes) and blocks installed apps (managed-app) from approving their own actions |