For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 725 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
Personal.ai deserves to be taken more seriously than almost everyone else in AI.
Not because its demos are flashier, or because its model is more capable, or because it found a better prompt wrapper around the same stateless assistant. The opposite. Personal.ai saw the thing the rest of the field kept stepping over. The AI product category was going to stop being about what a model can answer in one session and start being about what a system can become across many sessions. The primitive was not chat. It was not search. It was not even context. It was memory.
Their current public language is unusually strong. On the homepage, Personal.ai distinguishes context from memory cleanly: context is what the AI has been told; memory is what the AI has lived; identity grows from memory over time. That is not ordinary SaaS copy. It is a correct theory of the category.
Most AI companies are still selling shallower objects. The frontier labs sell capability. Enterprise search companies sell access to company knowledge. Email agents sell relief from a channel. AI-native OS projects sell a general-purpose environment. Personal.ai is selling accumulated identity.
That makes their current shape much less confusing. The consumer persona, Memory Stack, Gmail and Outlook integrations, AI message APIs, public Profile Page, enterprise Workspace, persona roles, channel directives, and carrier-native Memory Core are not unrelated product bets. They are expressions of the same thesis at different buying centers. If memory creates identity, then the winning product is wherever memory can accumulate, be governed, be invoked, and be paid for.
At the personal layer, the system lets a user train personas from documents, messages, integrations, and direct correction. The docs describe distinct personas with segregated memory, purpose, style, sharing settings, Autopilot and Copilot modes, and a Personal Score that measures how well an answer reflects the user's knowledge and style. The Memory Stack is the important object: a growing collection of memory blocks from uploads, messages, documents, and connected accounts. More memory makes the AI more tailored.
At the public layer, the Profile Page turns personas into surfaces other people can chat with. This matters because a personal AI is not only a private assistant. It can become a representative. A creator, executive, expert, or organization can expose selected personas to outsiders, making memory into an interface.
At the enterprise layer, the same object becomes a workspace. Personal.ai's enterprise launch is not just admin controls around chatbots. It is a governed persona factory. Teams can create, train, share, publish, and monitor AI personas. Persona templates automate the work of turning a role into instructions, recommended uploads, tasks, and expected outcomes. Persona owners, managers, and collaborators define who can train, manage, clear memory, delete, publish, or monitor. The release says the quiet part directly: personas can represent job functions, people, and organizations. That is a large claim. It says memory can make roles portable.
At the carrier layer, the thesis becomes infrastructure. Personal.ai's 2026 carrier page says an evolving AI can be bonded to every line on a network, with tokens treated as a fourth primitive beside talk, text, and data. Their AI Grid writing frames memory-based small language models near the edge as faster, cheaper, and more reliable than routing everything through centralized cloud LLMs. This is much more ambitious than a chatbot that knows me. It is a proposal that the telecom line becomes an AI identity rail.
That is why Personal.ai is one of the few companies in the category with a philosophy and a business model that actually rhyme. If identity emerges from memory, then every customer relationship is valuable to the extent that it grows a memory stack. The better the memory, the better the persona. The better the persona, the more usage. The more usage, the more memory.
The competitor map makes the point sharper.
Glean is closest in enterprise. Its knowledge graph and enterprise graph connect content, people, activity, permissions, projects, processes, and work execution. It may be the most obvious procurement choice for company knowledge. But Glean's center is the company graph. It asks how the organization can know and act through its corpus. Personal.ai asks how a memory-bearing persona persists and speaks.
Matrix OS is closest on altitude. It puts AI at the level of the whole computer: generated software, files as state, any-channel delivery, self-healing systems. That is the right height for the next computing interface. But its root metaphor is operating system: kernel, processes, files, system calls. Personal.ai's root metaphor is lived memory. Matrix OS asks how software becomes fluid around a person. Personal.ai asks how an AI identity grows from traces.
Second Me is closest philosophically. It is open-source, privacy-oriented, explicitly about an AI self that preserves and represents the user. It may be closer in soul to the future than most commercial products. But it does not yet show Personal.ai's commercial machinery: enterprise governance, public persona surfaces, API memory infrastructure, compliance posture, carrier economics, and edge deployment.
Cora is closest in felt pain. Give it the inbox; get your life back. It drafts in your voice, filters importance, and briefs the rest. That is a beautiful wedge. But the object is the inbox. Personal.ai is building the place where many wedges could accumulate into identity.
The frontier labs are closest in distribution and capability, but not in object. They have models, users, capital, and increasingly useful memory features. Still, memory remains subordinate to the assistant. For Personal.ai, memory is not a retention feature. It is the product, the moat, and the source of identity.
Hamilton Helmer's test helps because the strategic question is not whether Personal.ai creates value. It does. The question is what prevents a competent competitor from arbitraging that value away.
The strongest present answer is switching cost. Every memory block, correction, upload, integration, persona directive, public chat, workspace role, and channel assignment makes the system more specific. The user does not merely lose data by leaving. He loses accumulated fit. The enterprise does not merely lose search. It loses governed roles. The carrier does not merely lose token traffic. It loses the chance to make identity-bound AI a native network primitive.
Scale economies may appear if memory-specialized models near the edge are genuinely cheaper and faster. Counter-positioning may appear because generic assistants, cloud LLMs, and enterprise-search incumbents are optimized around different owned objects. Process power may appear if Personal.ai's memory lifecycle - encoding, stabilizing, storing, retrieving, updating - becomes hard to reproduce. Branding is already unusually strong. "Personal AI" is the category name everyone else wishes they had.
This is why the critique has to be careful. Personal.ai is not wrong in the normal way AI companies are wrong. It is right enough to make the next question unavoidable.
Who is the memory for?
For the person, memory means continuity. The AI remembers what matters and stops treating each conversation like a first date. It can speak in a known style, recall a history, carry a relationship, and improve through correction.
For the enterprise, memory means reusable expertise. The company can turn local knowledge and role behavior into governed personas that outlive individual availability and scale across channels.
For the carrier, memory means a new primitive on the bill. The line is no longer only connectivity. It becomes an identity endpoint whose token traffic can be bundled, metered, and monetized.
For the platform, memory means retention. The deeper the stack, the more valuable the system and the harder it is to leave.
All four can be true. The hard cases begin when they diverge.
Should a user be able to remove memory in a way that makes the AI worse but the person freer? Should an employee be able to veto a persona trained partly on their judgment if the company paid for the workspace? Should a carrier-native AI ever optimize for fewer tokens because the healthiest answer is less mediation? Should a platform make portability strong enough to weaken its own switching cost? Should Personal Score measure fidelity to the person, the persona, the workspace directive, or the current memory stack? Who decides that a trace was lived by the AI rather than merely observed by the system?
These are not gotchas. They are the questions Personal.ai earns by being serious.
The weak version of personal AI never has to answer them because it does not remember enough for the answers to matter. Personal.ai may remember enough. That is why it matters.
The company has found one of the central truths of the next computing era: context is not enough; memory is the path to identity. The next truth is less comfortable. Memory does not become personal because it is attached to an identity. Memory becomes personal only when the person holds the right to decide what the identity may keep, refuse, forget, and carry away.