The future of memory

1945

The memex A desk that remembers.

In July 1945 Vannevar Bush, who had run American wartime research, published “As We May Think” in The Atlantic Monthly.¹ He worried that knowledge was growing faster than anyone could use it, and he proposed a device for one person: the memex, a desk in which you would keep your books, records and correspondence on microfilm, and consult them at speed.²

The part that lasted was not the microfilm. It was the trail: a chain of items you had tied together yourself, which you could keep, retrace and hand to someone else. Bush’s memex was personal and associative. It held what you had read and written, and it followed how you thought.

1952 – 1968

Augmenting the intellect Cards, symbiosis, a demo.

Some of it was paper. From 1952 the sociologist Niklas Luhmann built a box of index cards, eventually some 90,000 of them, and wrote an essay about communicating with it, as if the box were a partner in his thinking.³

In 1960 J. C. R. Licklider described “man-computer symbiosis”: people would set the goals and make the judgments, and machines would do the routine work in between, so that the two together would think better than either alone.⁴ Two years later Douglas Engelbart wrote the framework for that partnership, a report called Augmenting Human Intellect.⁵

On 9 December 1968, in San Francisco, Engelbart showed it working: windows, hypertext, the mouse, linked files, revision control, video conferencing and real-time shared editing, in one ninety-minute demonstration.⁶ Almost every tool for thought since has been a piece of that afternoon.

1965 – 1972

Links, and a computer of your own Where things came from, and who held them.

In 1965 Ted Nelson coined the words hypertext and hypermedia.⁷ His Project Xanadu, begun in 1960, asked for more than links: documents built from pieces of other documents, links visible and followable from both ends, and every address naming its author.⁸ It was, among other things, an early demand for provenance: that a quotation should never lose the place it came from.

In 1972 Alan Kay described the Dynabook, a personal computer small enough to carry and meant for anyone, children first.⁹ The machine that holds your work would be yours.

1989 – 2001

The web, search, and a life in bits Everything linked; nothing yours.

At CERN in March 1989, Tim Berners-Lee proposed a linked information system for a lab where people came and went, and knowledge left with them.

“When two years is a typical length of stay, information is constantly being lost.”

The web he built linked everyone’s documents. Search made them findable: Larry Page and Sergey Brin’s PageRank, described in 1998, ranked a page by the pages that pointed to it.¹¹ In 2001 Microsoft Research began MyLifeBits, a lifetime store of everything in one man’s life, explicitly after Bush’s memex.¹² Personal knowledge management became a practice of its own: notes, tags, links, a second brain.¹³

Each of these made memory larger. None of them made it answer. You still had to know what you were looking for, and where you had put it.

2007

The attention era More to remember, less time to.

On 9 January 2007 Apple announced the iPhone.¹⁴ Within a few years mail, calendars, chat, documents and code lived in a dozen services each, all in your pocket, all asking for you.

“Hence a wealth of information creates a poverty of attention.”

Simon wrote it before any of this. The more there is to remember, the less of it anyone can hold, and the more a person’s context ends up scattered across services that do not speak to each other.

2017 – 2022

Machines that read And forget you between conversations.

In 2017 a paper titled “Attention Is All You Need” introduced the Transformer.¹⁶ Five years later, on 30 November 2022, ChatGPT put a large language model in front of everyone.¹⁷ For the first time, the machine could read what you had written and answer in kind.

But a model knows the world, not you. Each conversation started from nothing: what you decided last week, who your colleagues are, what you already ruled out. You told it again.

2020 – 2023

Retrieval, tools, agents Give the model what it needs, when it needs it.

Researchers answered with retrieval. Retrieval-augmented generation, in 2020, let a model look things up before it answered.¹⁸ ReAct, in 2022, had models alternate between reasoning and acting.¹⁹ Toolformer, in 2023, showed a model learning when to call a tool.²⁰ Generative agents kept a stream of memories and reflected on them,²¹ and MemGPT treated a model’s context like an operating system’s memory, paging in what mattered.²²

More context is not automatically better. Models use information at the start and end of a long input better than what sits in the middle.²⁸ What helps is the right context: small, relevant, and clear about where it came from.

25 November 2024

A protocol for context One plug, every assistant.

Anthropic released the Model Context Protocol, an open standard for connecting AI assistants to the systems where data lives, describing it as

“replacing fragmented integrations with a single protocol”

Within a year the protocol had a dated specification,²⁴ OpenAI’s Agents SDK spoke it,²⁵ and Google shipped Gemini CLI with built-in support for it.²⁶ For the first time, one source of context could plug into many assistants from many vendors.

2025

Every assistant, its own memory Context in fragments.

The assistants learned to remember. ChatGPT gained an optional memory and can recall past conversations.¹⁷ In September 2025 Anthropic brought memory to Claude, a separate memory for each project, and said it could be exported or brought in from another tool.²⁷

Each memory is useful inside its own walls. But people now use several assistants at once: one in the editor, one in the terminal, one in the browser, one on the phone. Each remembers a different fragment, in its own format, under its own rules. What you told one, the others never heard. So you repeat yourself. And none of them can show you what it knows about you, where each piece came from, or what it handed on.

Now

Whose words, and who decides Privacy and provenance stop being optional.

An assistant that reads your mail also reads what other people wrote to you. In 2022 Simon Willison named the attack this invites, prompt injection;²⁹ in 2023 researchers showed it working indirectly, through content a model retrieves.³⁰ In 2025 he named the dangerous combination plainly: private data, untrusted content and a way to send data out, in one agent.³¹

The answers point the same way. Local-first software argued that you should own your data and keep it on your own devices, with the network optional.³² The law gives Europeans the right to take their data with them.³³ Solid proposed that people keep their data in stores they control and grant access to it.³⁴ Media is getting signed records of where it came from.³⁵ Memory needs the same: whose words a line is, where it came from, and who was allowed to hear it.

Our view

Why a memory you own, beside every AI, is inevitable

Models will keep getting better, and there will keep being more of them. You will not use one. The thing that should not be rented, copied into each one, or lost when you switch, is what you have said and decided.

Your memory should belong to you, sit beside every assistant you use, and say only what you allow.

For that to work, it has to do a few things no single assistant can do for you:

1. Live where you are.On your own machine, encrypted, not in another company’s cloud.
2. Serve every assistant.Over an open protocol, the command line and hooks, not one vendor’s feature.
3. Say whose words each line is.Yours, someone else’s, a record, or a conclusion, never blurred together.
4. Let you decide what each one hears.Per assistant, with rules for names and topics, and parts that are never said.
5. Leave a receipt.What was sent, to whom, and what was held back, every time.
6. Keep the agent on its side of the line.An agent may read and narrow. Widening, spending and erasing are yours.

Bush’s memex was a desk for one person. The version for now is a memory for one person, that every machine they work with can consult, and none of them owns.

Today

Dory today What exists, and what it does not promise.

Dory is our first step. It runs on your Mac: it reads the sources you connect, remembers what you said and who said what, and serves a governed context capsule to your own AI tools over MCP, the command line and hooks. Every line says whose words it is, every assistant gets its own share, and every delivery leaves a receipt. An agent can set it up; anything that widens what Dory says waits for you and Touch ID.

It is not another assistant. If you could replace your assistant with Dory, we would have built the wrong thing: Dory is meant to be noticed only as your assistants knowing you.

It has limits, and we publish them. Instructions written as plain prose in other people’s messages can get through, and the slowest supported Mac has not been measured yet. The known limits say the rest. There is no public download yet; it opens at launch.

The MCP server The command Download

Vision

Beyond today Ideas, labelled as ideas.

Mnemo Technologies Labs, the trading name of Jacobo Mesonero Romanos, makes Dory. Our north star is short: own the memory, rent the intelligence. Models are rented and interchangeable; your context is yours. Everything below is a direction from our own plans, not a product. Where it is built, we say so; where it is only a name, we say that too.

Whatever comes next keeps the same rule as today: an assistant should know you, and you should know what it was told.