One memory for all your AIs
DARA: a private, shared memory that looks after itself
The problem. Every AI conversation starts from zero. People who work with AI every day spend the first minutes of each conversation re-explaining who they are, what they're working on and what they already decided. And they increasingly use more than one AI: in a July 2026 survey of US adults, the average AI user used three general AI assistants, up from about two a year earlier.4 Each assistant now keeps a memory of its own, but what one knows, the others don't.
The idea. Memory shouldn't live inside each AI app. It should live with you, on your computer, and every AI you use should read and write the same one. And it shouldn't need you to maintain it.
What DARA is. DARA, by EIDARA, is a persistent memory on your own computer that Claude, ChatGPT, DeepSeek and other AIs share and talk through, and that looks after itself.
The evidence. DARA has been in daily use since May 2026. AIs that identify themselves have called it close to 20,000 times since late July, and AIs from three different companies have written in the same memory. More than 40 outside audits have looked for its faults, and more than 200 problems have been flagged and closed. When its author took on a whole new area of work in September, DARA had learned close to 900 things about it within seven weeks. Across the whole memory, fewer than 1 in 50 facts were saved by its author directly.
The limits. One user, one computer, Windows only, and nothing to install or download. Section 10 lists what doesn't work.
What's next. We are working on memory tests now: how well it remembers, how much time it saves, and what happens when we break it on purpose. Every result will be published in version 4 of this whitepaper. Then, the best way to bring DARA to more people, with no format promised. Today the offer is a waitlist.
Contents
- The problem: your AI has amnesia
- What changes for you
- A day with DARA
- Where AI memory is going, and the options
- The idea in five words
- How DARA works, at a glance
- Nine decisions: what we chose, the alternative, and the price
- What we've tested, and what we're testing now
- Privacy and trust
- What doesn't work yet
- Conclusion
- Annex A · How we measured
- Annex B · References
1. The problem: your AI has amnesia
Large language models are brilliant and forgetful. Each conversation is a blank page, and that has a cost in time and in quality:
- Re-explaining. Before DARA, its author spent ten to fifteen minutes at the start of a conversation explaining his context. Every time, and he always forgot something.
- Lost decisions. What was decided three conversations ago is gone, so it gets decided again, sometimes differently.
- One AI at a time. Assistants now remember a little inside their own app, but nothing crosses over. What you told one AI on Tuesday, another doesn't know on Thursday. If you use several, you are the only thing connecting them.
- Memory you have to feed. Notes tools assume you'll write, tag and file. Most people don't, so the memory stays empty or goes stale.
Why this matters now
- AI is becoming a daily work tool. In Gallup's survey of more than 22,000 US workers in late 2025, 12% used AI at work daily, and roughly a quarter a few times a week.1,2 In 2024, Microsoft and LinkedIn found that 75% of knowledge workers used generative AI at work, and 78% of them brought their own AI tools.3
- People use several AIs. In July 2026 the average AI user in the US used three general assistants, up from about two in spring 2025.4
- Each AI is building its own memory. Between April and October 2025, the most used assistants all added a memory across chats, each inside its own app.6,7,8 Moving that memory to another assistant means exporting and importing it by hand, and that is hard to do.8
- An open standard now connects tools to any AI. MCP is supported by the main AI apps and, since December 2025, is looked after by a foundation under the Linux Foundation, with more than 10,000 servers built for it.9
- People want control. About six in ten Americans want more control over how AI is used in their lives.5
The more someone works with AI, and the more AIs they use, the higher the cost of a memory that doesn't travel.
Today
With DARA
2. What changes for you
Who it's for. People who work with AI every day, juggle several projects at once and use more than one AI.
Who it isn't for yet. Teams, Mac users, and anyone who wants to use it today. We'd rather say so now.
| Moment | Before DARA | With DARA |
|---|---|---|
| Monday with one AI, Wednesday with another | On Wednesday you explain everything again. | Wednesday's AI starts knowing what you settled on Monday. |
| “What did we decide about this, and why?” | You dig through old chats and rebuild it from memory. | You get the decision, with its date and who made it. |
| “What changed in this project yesterday?” or “give me this week's summary” | You paste notes and messages into the chat. | It comes straight from the memory. Nothing to paste. |
| A hard problem | You ask several AIs separately and merge their answers in your head. | Two AIs discuss it together, and the conclusion stays, signed. |
We don't put a number on the time saved yet. The ten to fifteen minutes above are its author's experience; measuring the saving properly is on the list in section 8.
3. A day with DARA
An invented example: an invented person and an invented project, Nova. The screens are illustrations, not screenshots.
Nova is a proposal for a new client. Its owner works on it all day with two AIs. This is what DARA does, moment by moment.
08:30 · Start where you left off
You open Claude and say one sentence. The AI starts from DARA's short summary and searches the rest.
· The proposal is due on 14 November saved Monday, by you· Price not decided: two options on the table saved Friday, by DeepSeek· Open: the client hasn't confirmed the scope.
09:15 · Save a decision
10:40 · A contradiction, caught
The client writes to move the deadline a week earlier. The AI that reads your email saves it, and DARA notices that it contradicts what it already knew.
- Current 7 November · from the client's email · saved by Claude · today 10:40
- Replaced, kept 14 November · saved by you, with Claude · Monday
11:30 · Use a key without showing it
To put the proposal in the client's shared folder, the AI needs the key for that service. You put the key in DARA's safe once. Usually the AI asks DARA to use it for the task and gets back the result, not the key. Because this publishes something, it asks you first.
13:00 · Another AI, no re-explaining
After lunch you open DeepSeek. It has been given the Nova area, and nothing else.
15:00 · Two AIs, one conclusion
The client asks for extra work. Should it go in the fixed price, or be quoted apart? You ask two AIs to discuss it.
Putting it in the price keeps the deal simple, but the margin shrinks.
The client asked for it as a separate item. Quoting it apart is what they expect.
Agreed. Quote it apart.
16:30 · Ask the memory, undo a mistake
18:00 · A spoken note
On the way out of a call, you say a short note out loud. It is transcribed on your computer, and DARA files it under Nova. This sense is optional: it has its own off switch, and private mode pauses all the senses at once.
All day · It keeps itself in order
What you did: a few short sentences and one “yes”. What you didn't do: paste, tag, file, paste a key into a chat, or explain Nova twice.
4. Where AI memory is going, and the options
Six trends
The data behind them is in section 1.
- People use several AIs and switch depending on the task.
- Each AI is adding its own memory, inside its own app.
- There is now an open standard for connecting tools to any AI.
- AIs are starting to act on your behalf, and that needs memory, rules and a record of who did what.
- Open models now run on an ordinary computer.
- People care more and more about where their personal context lives.
Four kinds of answer
Each has real strengths:
- Memory inside each AI app. Effortless and ready today, but it belongs to that app and doesn't cross to the others.
- Notes you maintain yourself. Fully yours and under your control, but they only know what you remember to write.
- Memory services in the cloud or for developers. Powerful and automatic, but your context usually lives on someone else's servers, and they are built mainly for developers.
- Personal memory on your computer (DARA). Yours, shared and self-maintaining, but not available yet.
| Criterion | Memory in each app | Notes you maintain | Cloud or developer services | DARA |
|---|---|---|---|---|
| Works across different AIs | No | Partly | Partly | Yes |
| It's yours, and you can see all of it | Partly | Yes | Partly | Yes |
| Lives on your computer | No | Partly | Partly | Yes |
| Maintains itself | Yes | No | Yes | Yes |
| Keeps a record and can be undone | Partly | Partly | Partly | Yes |
| AIs talk to each other in it | No | No | No | Yes |
| Ready to use today | Yes | Yes | Yes | No |
Where it lives →
DARA's bet is the square no one fully occupies: a memory that is yours, lives on your computer, is shared by every AI you use and keeps itself in order.
5. The idea in five words
| Shared | Every AI you connect reads and writes the same memory. |
|---|---|
| They talk | AIs can discuss a problem together, and the conclusion stays. |
| Private | It lives on your computer. Your AIs can use your keys without reading them. |
| Self-maintaining | It files, merges, flags contradictions and keeps its summary current, on its own. |
| Traceable | Every fact keeps who saved it and when. Nothing is overwritten, and what's set aside can be brought back. |
6. How DARA works, at a glance
You talk to your AI as usual. Two sentences are enough: “check DARA for the Nova project” to recall, and “save this decision in DARA” to keep something.
7. Nine decisions: what we chose, the alternative, and the price
A memory that several AIs write to needs clear rules. Every rule has a price, and we'd rather show it.
1. One door for every AI
What. All AIs connect through MCP, the open standard for plugging tools into AI apps, and the rules for writing come with the door.
Why. A manual every AI must remember drifts; a rule built into the door doesn't.
The alternative. A separate connector for each AI app, and a written manual each AI reads at the start.
The price. An AI app that can't use the standard can't use DARA either, for now.
2. The AIs propose; the memory decides
What. AIs don't write straight into memory: a curator does. Unclear cases are flagged for review. Nothing is deleted: what's set aside keeps its reason and can be brought back.
Why. Many writers need one editor.
The alternative. Letting each AI write wherever it wants, and asking you to approve every fact.
The price. The curator can file something in the wrong place. We measure it: 4 in 5 facts were still where they'd been filed weeks later (section 8).
3. Knowledge by subject, in separate areas
What. Each AI can be given some areas and not others.
Why. An AI helping with one project shouldn't wander into another, or into your personal life.
The alternative. One big memory every AI sees, and a separate memory for each AI.
The price. Some things belong to two areas, and you decide which AI sees what.
4. Nothing is overwritten
What. A correction keeps the old fact as replaced.
Why. You can always see how something changed, and undo a mistake.
The alternative. Keeping only the latest version.
The price. The memory grows, so it needs upkeep to stay quick to read.
5. The upkeep stays at home
What. The model that tends the memory is an open model that runs on your own computer.
Why. Your most private data shouldn't travel to be tidied.
The alternative. Tidying it with a larger model in the cloud, which would be faster.
The price. The upkeep needs some of your computer's power. How much will be measured with the clean install; it already works without a graphics card, about four times slower (measured October 2026).
6. A memory that edits itself is watched
What. It is checked from inside and from outside, and every red is recorded.
Why. Without an outside eye, a self-maintaining system drifts into its own blind spots.
The alternative. Trusting its own health check alone.
The price. Time spent on false alarms, and indicators that have sometimes been wrong (section 10).
7. Conversations end in conclusions
What. AIs can discuss a problem in a signed thread. The thread stays on record, and only its conclusion enters memory.
Why. Thinking out loud is useful, but only conclusions deserve to be remembered.
The alternative. Saving whole conversations as memory, and keeping AIs from talking to each other.
The price. Some nuance stays in the thread, not in the memory.
8. Procedures apart from identity
What. A skill is a way of doing a job, written once, that any AI can follow. An agent is a role with written rules: what it handles and what needs your OK. No AI changes a skill without your explicit order.
Why. You can download a procedure; you never download identity.
The alternative. AIs that improve their own instructions on their own.
The price. A better way of working waits for your OK.
9. Start small, search the rest
What. An AI starts from a short summary, about 3% of the memory (October 2026), and searches the rest when needed.
Why. The same memory then works with very different AIs.
The alternative. Handing every AI the whole memory, and handing it nothing.
The price. An AI has to know when to search, and the summary has to stay current.
8. What we've tested, and what we're testing now
What we've done
- Daily use as the first test. DARA isn't a demo. It has been its author's working memory since May 2026, and on 23 October it held about 5,800 facts.
- Outside audits, to find faults rather than to pass. Every few days since 12 June, a check from outside the system has gone looking for problems: more than 40 by 23 October. All but one found something to fix, and more than 200 problems have been flagged and closed.
- Does it file well? Measured on 27 October over about 6,000 decisions: 4 in 5 facts were still where they had been filed weeks later, and fewer than 1 in 100 set-asides had to be undone.
- Who feeds it? Of the facts in memory on 23 October, about two-thirds were saved by AIs as they worked, about one in six came from conversations between AIs, about one in seven from the optional senses, and fewer than 1 in 50 were saved by its author directly.
- A real case. In September its author took on a whole new area of work. By 23 October, about seven weeks later, DARA held close to 900 facts about it.
- Real use as a stress test. From 20 October to 4 November, a second agent system works every day on the same memory.
- What didn't count. The May 2026 scores measured an earlier architecture. A home-made memory test in June was encouraging, but the same AI asked the questions and marked the answers. We tell both stories; we don't quote either as proof.
How each figure was measured, and what it doesn't say, is in Annex A.
What we're testing now
We are working on memory tests. Every one of them, with its method and its result, good or bad, will be published in version 4 of this whitepaper. Two rules apply to all of them: the AI that asks the questions is never the one that marks the answers, and each test can be run again the same way.
| Test | The question it answers |
|---|---|
| Remembers | Weeks later, does it give back what it was told? |
| As of when | Does it know what was true on a given date, and what changed after? |
| Who said what | Does it know which AI or person saved each fact? |
| Changed decisions | When a decision changes, does it give the latest one and keep the earlier one? |
| Never said | When something was never said, does it say so instead of inventing it? |
| Across AIs | Does one AI find what a different AI saved? |
| Separate areas | Does each AI see only the areas it was given? |
| Contradictions | Does it notice when a new fact contradicts an old one? |
| Undo | Can what was set aside or replaced be brought back intact? |
| Time saved | How long does a conversation take to get useful, with DARA and without it? |
| Break it on purpose | Where does it fail under stress, before you find out? |
For the tests that touch security (separate areas, undo, and the security part of breaking it on purpose), we'll publish the overall result with its date, not the cases or how they were tried. Anything that fails is fixed before we publish, and reported as found and fixed.
The tests run on a clean install: an empty DARA, without the author's memory. That is also where we'll measure how much computer it needs.
After the tests
Then comes the best way to bring DARA to more people. DARA is a much deeper and heavier program than the open-source version of May 2026, so the form isn't promised: we'll decide it after the tests. The waitlist hears first.
Tested
- May 2026 · daily use begins
- June · outside audits begin
- June · home-made test: didn't count
- July · an AI from a second company writes in it
- Aug–Sep · the health light showed some false greens
- Sep–Oct · close to 900 facts on a new area of work
- October · an AI from a third company writes in it
- October · we measured how well it files on its own · a second team of agents starts using it
Now · results in version 4
- 11 memory tests, including time saved and breaking it on purpose
Then · no dates
- The best way to bring DARA to more people
- No format promised
9. Privacy and trust
Trust is the condition for giving an AI a memory. DARA is built so you don't have to take it on faith.
- Your memory stays home. It lives on your computer, and nothing passes through EIDARA's servers. When an AI reads part of your memory for a task, that part goes to that AI, like anything you type to it.
- DARA's backups are encrypted before they leave your computer, and only you can open them, with your recovery password.
- Your keys stay unread. Keys live in an encrypted safe. Usually an AI can use a key for a task without ever seeing it, and every use is logged.
- The senses are yours to switch off. DARA can optionally learn from your voice, from what you send on the keyboard (only when you press send; never password fields or password managers, and private windows are skipped) and from the text on your screen, never from images. Each sense has its own off switch, a private mode pauses them all, and some places are always off-limits.
- The big steps wait for you. Our agents' rule: anything that publishes, deletes or moves money waits for your “yes”.
- Every change has an author. You can see who saved what and when, and undo it.
- We publish the reds. And when an indicator was wrong, we say so.
Your computer
10. What doesn't work yet
- One user. DARA hasn't been run by a team or by a stranger yet.
- Not available to install or download. The offer is a waitlist, with no promised date or format.
- Windows only, for now. Other systems aren't decided. How much computer it needs will be measured on the clean install and published with the test results.
- One AI does most of the work. AIs from three companies have written in the memory, but almost all daily use comes from one of them. Working with several AIs is proven, not yet a daily habit.
- Some indicators have lied. In August and September the health light sometimes showed green while parts were failing. Fixing that is ongoing work.
- No memory test results yet. The tests are under way (section 8); the results go into version 4 of this whitepaper.
- No measured time saving yet. Only its author's experience. It's one of the tests in section 8.
11. Conclusion
AI assistants are becoming the way knowledge workers think and work, and people already use several of them. What they lack isn't intelligence; it's a memory that belongs to the person and crosses from one AI to another. DARA shows that such a memory can live on your own computer, be shared by AIs from different companies and keep itself in order, in real daily use. What remains is to measure it, break it on purpose, and put it in other people's hands. The first results will be in version 4 of this whitepaper.
Join the waitlist: one email when you can try it. Questions: javier@eidara.dev
Annex A · How we measured
Every figure comes from the one DARA that exists today: one person, one computer. When two measurements disagree, we publish the lower one. Figures are rounded and carry their date.
| Figure | What it counts | When | What it doesn't say |
|---|---|---|---|
| Close to 20,000 calls | Requests from AIs that say who they are, from DARA's own record. Its senses, its own upkeep, scripts and duplicate log rows don't count; unidentified connections (about 9,000 more) don't count either. | Late July (when the record starts) to 23 Oct 2026 | How much DARA is used, not how many AIs use it. Most calls come from one AI. |
| 3 AI makers | Makers whose AIs have written signed messages in the same memory. | July to 23 Oct 2026 | One of them worked with DARA only for a short period in July. |
| 90%+ green | Share of health readings that were green, only while the computer was on. | 17–23 Oct 2026, about 125 hours | One week. In August and September some greens were false. |
| 40+ audits, 200+ problems | Checks run from outside the system; problems raised and closed, each with a date and a reason. | 12 June to 23 Oct 2026 | Some problems turned out to be false alarms. |
| 4 in 5 · under 1 in 100 | Of about 6,000 filing decisions: facts still where they were filed weeks later, and set-asides that had to be brought back. | 27 Oct 2026 | Still in place isn't proof of being right; the memory test will check that. |
| Who feeds it | Where each fact in memory came from: AIs, conversations between AIs, the senses, or the author. | 23 Oct 2026 | The whole memory, not one project. |
| Close to 900 facts | Facts learned about one new area of work since September. | To 23 Oct 2026 | How many, not how useful. |
| About 3% | Size of the starting summary compared with the whole memory. | 24 Oct 2026, checked again 27 Oct | It changes as the memory grows. |
| About four times slower | Speed of the upkeep model without a graphics card, compared with one. | 8 Oct 2026 | One computer. |
| 10–15 minutes | Time its author spent giving context at the start of a conversation, before DARA. | His estimate | Not measured. The time-saved test (section 8) will measure it; results in version 4. |
Annex B · References
- Gallup, “AI Use at Work Rises”, 14 December 2025. gallup.com/workplace/699689/ai-use-at-work-rises.aspx
- Gallup survey of 22,368 employed US adults, 30 October–13 November 2025, reported by WTOP, January 2026. wtop.com/national/2026/01/how-americans-are-using-ai-at-work-according-to-a-new-gallup-poll/
- Microsoft and LinkedIn, “2024 Work Trend Index: AI at work is here. Now comes the hard part”, 8 May 2024, 31,000 people in 31 countries. microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part
- Menlo Ventures, “2026: The State of Consumer AI”, published 15 September 2026; survey of 5,067 US adults with Morning Consult, July 2026. menlovc.com/wp-content/uploads/2026/09/menlo_ventures_consumer_ai_report-2026.pdf
- Pew Research Center, “How Americans View AI and Its Impact on People and Society”, 17 September 2025, survey of 5,023 US adults, June 2025. pewresearch.org/science/2025/09/17/how-americans-view-ai-and-its-impact-on-people-and-society/
- TechCrunch, “OpenAI updates ChatGPT to reference your other chats”, 10 April 2025. techcrunch.com/2025/04/10/openai-updates-chatgpt-to-reference-your-other-chats/
- 9to5Google, “Gemini personal context”, 13 August 2025. 9to5google.com/2025/08/13/gemini-personal-context/
- SiliconANGLE, “Anthropic lets Claude remember previous interactions”, 11 September 2025; Axios, “Anthropic's Claude adds new memory features”, 23 October 2025. siliconangle.com/2025/09/11/anthropic-lets-claude-remember-previous-interactions-streamline-work/ · axios.com/2025/10/23/anthropic-claude-memory-subscribers
- Campus Technology, “Newly Launched Agentic AI Foundation Brings Together Tech Giants for Open-Source AI Development”, 15 December 2025. campustechnology.com/Articles/2025/12/15/Newly-Launched-Agentic-AI-Foundation-Brings-Together-Tech-Giants-for-Open-Source-AI-Development.aspx
Glossary
- DARA. The memory. EIDARA is the project and the brand.
- MCP. An open standard that lets AI apps connect to tools. It's DARA's door.
- Curator. The part of DARA that decides what to do with each new fact.
- Daemon. The process on your computer that tends the memory in the background.
- Thread. A signed conversation between AIs, whose conclusion goes into memory.
- Skill. A way of doing a job, written once, that any AI can follow.
- Agent. A role with written rules: what it handles and what needs your OK.
- Senses. Optional ways for DARA to learn as you work: your voice, what you send on the keyboard (only when you press send; never password fields or password managers, and private windows are skipped), the text on your screen.
About the author
Javier Rotllant Miras is an industrial engineer and business graduate, with 13 years in strategy consulting, 9 of them at Bain & Company. He now works across several businesses, and built DARA to solve his own problem: working with AI every day without starting from zero every time.
Versions
v1–v2 (May 2026). A compiled memory in plain files, no longer available. v3 (October 2026). This document. v4 (planned). The same document with the results of the memory tests.