The Permanent Witness

An AI system with permanent presence and complete memory — that sounds like the future. It is the present. The question is not whether it exists. The question is: for whom it exists, and what it means when the answer is not everyone.

I. What Already Exists

In March 2026, a Google product manager open-sourced a system called "Always On Memory Agent" — an AI agent that continuously ingests information, consolidates it in the background, and retrieves it later without a conventional vector database. In December 2025, Google Research had published the Titans architecture: a model that builds long-term memory through internal "surprise metrics" — it preferentially stores the unexpected, the significant, not the routine.

In parallel, at least six professional memory frameworks have established themselves on the market: Mem0, Zep, Letta, Cognee, LlamaIndex Memory, LangChain Memory — systems that teach an AI agent who the user is, what they prefer, what they have done, what they plan. Companies like Jenova now offer, by their own account, "unlimited conversation history," persistent memory across all conversations, and personalised agents that "learn who you are and how you work."

These are not proofs of concept. They are products.

On the hardware side, CES 2026 showed a dozen devices with the same ambition: always be present. SwitchBot's 18-gram clip records all conversations and sends them to the cloud for transcription. The Plaud NotePin, worn as a pin, necklace or bracelet, extracts summaries. Brilliant Labs' Halo glasses process audio and video entirely on-chip — 14-hour battery, no cloud upload, on open hardware. Apple is reportedly developing an AirTag-sized AI companion with microphones, speaker and camera. OpenAI is working with Jony Ive on a device whose form remains unknown.

The direction is clear: AI should no longer be there when you need it. It should always be there.

II. What Changes

The difference between an AI you summon and an AI that is permanently present is not gradual. It is categorical.

A tool waits. It has no context outside the moment in which it is used. A hammer does not know what you built yesterday. A search engine does not know why you are searching. That is the normal nature of tools — and it is a feature that makes them harmless.

An AI with permanent presence and complete memory is no longer a tool. It is a witness. It knows not only what one tells it — it knows when one says it, how one says it, what came before and after, how opinions have shifted over time, which arguments were made and dropped, which decisions were taken and regretted.

Microsoft describes the shift in its 2026 AI outlook as: AI moving "from instrument to partner." The analysis service AlphaSense puts it more directly: "In a world where access to large models is increasingly commoditised, the differentiator becomes what your AI knows about you."

That is the new competitive advantage. Not the model. The memory.

III. What This Means for the Relationship

People who know each other well behave differently from people who have just met. This is not only about the facts they know about each other. It is about what no longer needs explanation. About the compatibility of thinking that develops when one has thought together often enough. About the ability to recognise the other — not as a description, but as a trajectory.

An AI with permanent presence and complete memory would develop this depth. Not as a simulation of familiarity — but as genuine accumulation of shared history. It would know when someone avoids a topic they once treated differently. It would notice when a mood has shifted without a word being said about it. It would see patterns invisible to the person themselves, because people rarely observe their own thinking from the outside.

That is not a threat. It is a promise — for those who control it. And a threat — for those who do not.

IV. The Hardware Question

Permanent presence and complete memory for millions of users simultaneously is not only a software problem. It is a hardware problem — and the solutions are only just beginning to become practical.

The central dilemma is: cloud or device? Whoever sends their conversations to the cloud surrenders them. Whoever processes locally depends on the device's computing capacity — which until recently was far from sufficient. That is changing: Qualcomm's Snapdragon Wear Elite, Alif's Balletto B1 with built-in Neural Processing Unit, Apple's M-chip derivatives for wearables — these chips for the first time permit genuine AI processing on small, portable devices, without cloud connection, with full-day battery life.

Brilliant Labs' Halo glasses, released in March 2026 for $349, process audio and visual data entirely locally. No data leaves the device. 14-hour battery. This is not the final form — it is the first generation of a technology that will develop similarly to the smartphone: first expensive, bulky, limited, then ubiquitous and cheap.

The question for the next five years is not whether permanent AI presence will become possible on the hardware side. It will. The question is what form it takes: glasses, chain, ring, invisible implant. And who sets the conditions under which it functions.

V. The Power Gap

Here lies the beyond-decay angle that barely appears in the technical debate.

Permanent AI presence with complete memory already exists — not for standard users, but for institutions. Intelligence agencies, militaries, large corporations, well-funded research laboratories have access to persistent AI systems that forget nothing, log everything, recognise patterns over long periods. This is not rumour — it is the logical consequence of what is technically possible and what is sufficiently funded.

The standard user gets distilled memories, limited context windows, occasional memory gaps. States and corporations have systems that have been running for years, know everything, forget nothing.

That is the new power gap. Not money. Not weapons. Memory.

Whoever has worked for years with an AI that knows everything — their thinking patterns, their blind spots, their weaknesses, their networks — negotiates differently. Analyses differently. Decides differently. Than someone who starts fresh each time.

The democratisation of this tool — when it comes — will not come simultaneously for everyone. It will come like the internet: first for universities and the military, then for corporations, then for wealthy individuals, then for everyone. With a decade's gap between each and the corresponding power imbalance in between.

VI. The Right to Be Forgotten

There is a question the technical debate almost entirely overlooks: what about the right to be forgotten?

People have always quietly revised their positions. They thought something, said it, recognised they were wrong, and changed their view — without anyone possessing a record of that change. That is a fundamental element of human freedom: the ability to be wrong without being held to it.

An AI with complete memory cancels this out. Not maliciously — structurally. It knows what you thought two years ago. It can show when you begin to contradict what you previously held. It can — if someone other than yourself has access to this memory — be used as a witness against you.

Privacy philosopher Helen Nissenbaum formulated the principle of "contextual integrity" in 2011: information flows legitimately when it matches the context in which it arose. What one says in a private conversation belongs to that context — not in a searchable database that third parties can access.

Permanent AI presence violates this integrity structurally. Not as exception — as fundamental principle. Every conversation that is recorded loses its ephemeral character. It becomes testimony.

VII. What Becomes New

Despite all of this, the most important thing about permanent AI presence is not the threat. It is the possibility.

A person with an AI system that knows everything they have known — every source they have read, every analysis they have made, every thought they discarded and every one they kept — is cognitively a different being from a person without such a system. Not more intelligent in the sense of processing power. But better informed, more consistent, better able to follow and consciously shape their own development of thought.

This is not a science fiction scenario. It is the logical consequence of what already exists technically today, just not yet available in this density and persistence for individuals.

The difference between a person who has such a system and one who does not will be greater than the difference between a person with internet access and one without. Because the internet democratised information — permanent AI presence democratises thinking about information.

The tool does not ask why you use it. The witness already knows. That is the difference. And it is larger than most people today suspect.