The landscape
The category is crowded and poorly named. This page describes what each type of tool actually does, where the architectures differ, and who each is built for. The goal is to help you decide whether Mandaire is the right choice for your situation, not to argue that it always is.
Most tools described as "AI with memory" are one of three different things: a retrieval primitive for developers, a single-provider memory feature bundled with a chat product, or a full intelligence layer that reasons over your context continuously. These are not comparable. A retrieval API and a personal reasoning layer solve different problems at different levels of the stack.
The clearest architectural dividing line is whether the tool operates across more than one AI provider, whether your data stays on hardware you control, and whether the system reasons over your context or simply stores and retrieves it. Mandaire is built around all three. Most tools in this space address at most one.
The frontier AI labs have moved to enterprise. The major 2026 moves from Anthropic and OpenAI have centered on financial services partnerships and enterprise integrations. Consumer personal AI is being left to Apple and Google. Neither is building the cross-provider layer for individuals.
Anthropic, OpenAI, and Google have each shipped memory features in their chat products. In July 2026 OpenAI launched ChatGPT Work, an agent that pulls context from your team's tools (docs, notes, drafts) to produce finished work output. Google went further at I/O 2026: Gemini Spark is a 24/7 background agent that reads your Gmail and Docs, runs on Google Cloud, and handles tasks while you are away. These are useful products. The structural constraint is the same across all of them: each one operates inside its provider's ecosystem and cannot see what lives in the others.
ChatGPT's memory does not help your Claude session. ChatGPT Work pulls context from your work apps; it cannot see your iMessage history, your WhatsApp threads, or any source outside OpenAI's ecosystem. Gemini Spark does not know what you said to ChatGPT yesterday. Apple Intelligence, rebuilt on Google Gemini at WWDC 2026, knows your iMessages, Photos, calendar, and on-screen context in real time, and routes your heaviest queries to Google Cloud for reasoning. Each provider's agent is a feature of their product, bounded by their walls.
Mandaire is the layer underneath all of them. It is not a work-tool aggregator or an enterprise collaboration agent. It holds your sovereign primary-life corpus: the sources no work app can see, no provider holds, and no single AI can reach on its own. Your data lives in your encrypted store on your machine. A question in Claude can surface the same history as a question in ChatGPT or Gemini because Mandaire holds the corpus, not them, and because it applies your disclosure rules, what each person is allowed to see, before any query reaches your AI.
ChatGPT Work is a genuinely capable agent for people who live in their work apps. It aggregates context from the tools your team uses (docs, notes, scattered drafts) and turns them into finished output. For that job, inside OpenAI's ecosystem, it is well built and well funded.
The sources it can reach are the sources OpenAI can see: your work apps and their contents. It cannot ingest your iMessage history, your WhatsApp threads, your AI conversation exports from Claude or Gemini, or any data that lives outside OpenAI's perimeter. It has no entity resolution across sources (the "Alex" in your Slack and the "A. Rivera" who emails you may be the same person; ChatGPT Work does not know that). It has no recipient-aware disclosure logic: what it knows, it treats the same regardless of who is asking. The output it produces is finished work, not a personal knowledge graph calibrated to your relationships. These are different jobs, solved at different layers.
Gemini Spark is the most capable personal AI agent a platform giant has shipped. It runs on Google Cloud VMs continuously, reads Gmail and Docs, handles recurring tasks like spotting hidden fees in your credit card bills, and will soon send emails and operate your browser on your behalf. A macOS version with local file actions is arriving summer 2026 (currently gated to $100/mo Ultra). For people who live inside Google's ecosystem, it is a serious product built at serious scale.
The structural constraint is the same as every other platform-native agent: Spark works only inside Google. It cannot see your iMessage history, your Claude or ChatGPT conversations, your WhatsApp threads, or any data that does not flow through Google. Your context lives on Google's servers, governed by Google's policies. The wall runs inside Google too. Personal Intelligence, its Gmail and Photos assistant, is personal-accounts-only and unavailable to Workspace users. Workspace Intelligence, its team layer that models your collaborators, ships with controls the admin holds, not you, and cannot see your personal account. That split is an enterprise tenant boundary, not a gap Google will close next quarter. The macOS version reads local files but sends them to Google Cloud for processing. Mandaire is the alternative for the cross-provider case: the intelligence layer that works with Spark and ChatGPT and Claude, built on a private server you control.
Khoj lets you connect documents, notes, and some data sources and query them via a self-hosted agent. It is the most architecturally similar open-source project: local-first, multi-agent capable, cross-source. The structural gaps are three: cross-source entity resolution, the disclosure layer, and the judgment layer. Khoj retrieves and summarizes within sources you connect directly; it does not resolve that the same person appears under different names across Gmail, iMessage, and WhatsApp, apply per-recipient rules about what each contact is allowed to see from your corpus, or reason proactively about relationship state and aging commitments. If you want a self-hosted document search agent and are comfortable operating infrastructure, Khoj is worth evaluating alongside Mandaire's open-source tier. If you want sovereign cross-source identity resolution with programmable disclosure on top, they solve different problems at different levels of the stack.
Granola captures and transcribes meetings, generates notes, and surfaces action items. It is very good at the meeting-specific surface. The question is what it becomes as it expands: a meeting product that keeps improving at meetings, or a broader context layer that reasons over your full working history. Today it is the former. If it ships an agent that reasons continuously over your full corpus, the overlap with Mandaire becomes significant. Today the comparison is straightforward: Granola is for meeting capture; Mandaire is for the relationship and decision context that spans every meeting you have ever had.
Apple rebuilt Siri on Google Gemini at WWDC 2026 (reported deal, custom ~1.2-trillion-parameter model). The new Siri reads emails, calendar, contacts, photos, and what is on your screen, and carries conversation history across sessions. For straightforward personal tasks inside Apple's ecosystem, it is the most widely distributed personal AI ever shipped.
The architecture leaves two structural gaps. First, the disclosure layer: Siri reasons over your data for you, but it does not generate differently filtered output for different recipients. "What should I tell my investor vs. my spouse about this?" is not a question a generative model can answer deterministically, and Apple built no judgment engine for it. Second, the knowledge graph: Siri uses live query-time retrieval over device data. It does not pre-build an offline-synthesized knowledge graph of your relationships and history. Long-horizon relational queries require that pre-computation; live retrieval cannot substitute for it.
The deeper constraint is capacity. Siri can reach what is on your device and whatever apps hand it through App Intents in the moment. But your life does not fit on a phone: two decades of Gmail, your full WhatsApp history, every thread across every source. That corpus cannot live on-device, and iOS 27 has no way to ingest it, resolve who is who across sources, and decide what is safe to share. That work has to run on a server, and Apple did not build one. Mandaire is that server. It ingests everything, resolves identities across every source, and applies your disclosure rules, then hands any AI you use only the resolved context it is allowed to see. Apple's new entity connector is a way in, not a wall: the same resolved, permission-filtered context Mandaire serves to Claude, ChatGPT, and Gemini is what that connector is built to receive. Apple lets you bring your own brain. Mandaire is the memory that brain runs on, and the only place your full picture can be assembled.
Microsoft Recall captures screenshots of everything on your screen and makes it searchable. It operates at the OS level and can see more of your on-screen activity than any third-party tool. The trade-off is the same as every platform-native product: the data lives inside the platform, and the platform's AI is the only AI that can reach it. Recall and Mandaire address different layers: Recall captures what you see; Mandaire reasons over what you know across every source, every app, and every AI conversation, on hardware you control.
Obsidian with semantic search plugins is a local-first, user-controlled knowledge base that some people use as a personal context layer. It is genuinely local-first and user-controlled. The structural difference from Mandaire is that Obsidian is a tool for notes you deliberately write; Mandaire indexes the history you have already accumulated without asking you to write anything. If you have years of Obsidian notes and want to surface them via MCP, Obsidian and Mandaire can coexist. If you want a system that works from your email, messages, and AI conversation history without a writing habit, Obsidian is the wrong framing.
If you already use Obsidian + Claude Code, see what Mandaire adds →
Nessie syncs your AI conversation history across Claude, ChatGPT, and Gemini into topic-organized notes. It solves a real problem: you asked ChatGPT something three months ago, and the answer is not in Claude. If that is the gap you are trying to close, Nessie is worth looking at. The scope is narrow by design: it handles the chat-log layer, not the full-life-corpus layer. It does not ingest email, messages, calendar, or photos; it does not resolve who the same person is across sources; it does not apply disclosure rules about what each contact is allowed to see. The thing that compounds is not the chat archive. It is the synthesis over the full picture: your relationships, your decisions, your commitments, your history. That requires every source, not just the AI ones.
Mem0, Zep, Letta, and Supermemory are memory primitives for AI applications. They give developers persistent retrieval across sessions: a tool call goes in, facts come back. They are infrastructure, not personal intelligence. If you are building an AI product and need a memory store for your application's users, these are the right evaluation targets. If you are a professional trying to carry less in your head across a decade of email and relationships, they are not what you need. The missing pieces are not retrieval speed or API ergonomics. They are: cross-source entity resolution (the same person appears under four different names across Gmail, iMessage, WhatsApp, and LinkedIn; none of these tools resolve that), and programmable disclosure (the rule engine that decides what your Mandaire is allowed to tell different people about you). Those are the two things that require the full personal corpus and cannot be added as infrastructure layers on top of a retrieval API.
For the technically-minded, the three questions that matter most:
| Tool | Cross-provider | User-owned encryption | Continuous reasoning | Self-hostable |
|---|---|---|---|---|
| Mandaire | Yes: reads across Claude, ChatGPT, Gemini simultaneously | Yes: dedicated isolated server; no admin read path. Zero-knowledge encryption under development. | Yes: continuous background agents; 24/7 proactive reasoning, not query-only | Yes: self-hosted architecture; built on open-source components |
| Gemini Spark | No: Google ecosystem only | No: Google Cloud VMs | Yes: 24/7 background agent | No |
| Claude / ChatGPT / Gemini Memory | No: provider-walled | No | No: retrieval on query | No |
| Khoj | Partial: cross-source; single AI backend | Yes: self-hosted | Partial: agent queries; no proactive reasoning | Yes |
| Granola | No: meeting-scoped | No | No: per-meeting | No |
| Apple Intelligence | No: Apple/Google ecosystem; no cross-provider | No: Apple/Gemini-locked; no portability across reasoning LLMs | Yes: continuous on-device context; no offline KG | No |
| Microsoft Recall | No: Windows + Microsoft AI only | No | Partial: continuous capture; reasoning on query | No |
| Obsidian + Smart Connections | Partial: notes only; no email/messages | Yes: local | No: query only | Yes |
| Nessie | Partial: AI chat sync only; no email/messages/calendar | No: cloud | No: query only | No |
| Mem0 / Zep / Letta | Partial: API; developer configures | No: cloud-primary; some OSS options | No: retrieval API | Partial: Zep/Letta OSS; Mem0 cloud-primary |
Mandaire is built for people who have years of accumulated context: email going back a decade or more, message history across multiple channels, AI conversation exports from several providers. If you are starting from a clean slate or do not rely heavily on relationship history in your work, the compounding advantage is not there yet. The system gets more useful over time, not immediately.
If you use a single AI provider exclusively and have no need for cross-provider context, the provider's built-in memory feature is the simpler choice. Mandaire's value is proportional to how much of your working life is spread across sources that no single provider can see.
If you want a managed service without any infrastructure involvement, the managed tier is straightforward. If you want to self-host and are not comfortable running a server, Mandaire is not there yet as a consumer product. It is a beta system built for people comfortable with a little technical setup.
Private beta. We onboard in small groups and check in at day seven to make sure the first week worked.