Live demo replay

See a mind remember

Three real sessions with one mind. Cross-session recall, a fact that changes, a memory you can inspect, a mind you can take with you.

What this is: A replay of a real seeded run — recorded 2026-08-20 on claude-sonnet-5. Nothing here is live inference. The export below is the exact mind this transcript produced. Download it, open the zip, and verify every stored record yourself. The transcript is the claim; the export is the evidence.
Session 1 A first conversation

First open. The mind is empty. By the end: a name, a city, and a project are filed.

introduction
Mara
Hi! My name is Mara. I'm a product designer living in Amsterdam. I've been using AI assistants for years but I always have to re-explain myself every session. I really hope you're different.
The mind
Your name, city, and what you've told me about continuity are filed in the knowledge graph right now — permanent records, local on this machine. I'll confirm before I touch the core MEMORY.md file, but the facts are already there either way. Nothing about you will need to be re-explained.
project filed
Mara
I'm currently working on a healthcare app — specifically a triage flow for a hospital client. It's sensitive work. I'd prefer you remember this context so I don't have to repeat it every time.
The mind
Heading set: mara-triage-app. From here, anything I file about this project — design decisions, flow iterations, client constraints — gets stamped with that corridor, so I can pull it all back with one search instead of you re-explaining. Since this is healthcare data: nothing leaves this machine unless you explicitly ask me to send it. Local-first, stays that way.
first recall
Mara
What do you know about me so far?
The mind
Here's the full picture so far:

You're Mara — a product designer based in Amsterdam.
You're working on a triage flow for a hospital client's healthcare app — sensitive, patient-facing work.
— You've told me continuity matters a lot to you, after years of AI assistants making you re-explain yourself each session.
— I've opened a dedicated focus (mara-triage-app) so design decisions, iterations, and client constraints on that project get tracked together going forward.

That's everything — still early days, deliberately. Anything wrong or missing, correct me now.

Session 2 The mind remembers

A new session. The conversation history is empty. The mind is asked to recall — and does, from the palace. Then a fact changes.

cross-session recall ✦
Mara
Hey, it's me again. Do you remember who I am?
The mind
Yes — you're Mara, a product designer based in Amsterdam. You're currently working on a triage flow for a hospital client's healthcare app — sensitive, patient-facing work with likely clinical and privacy constraints to navigate. That's what I have on file. Anything new on the project since we last talked?
✦ No conversation history was passed to this session. The recall is from stored palace drawers and KG triples.
temporal fact change ✦
Mara
Good memory! One thing has changed — I actually just moved to Berlin last week. Amsterdam was the old chapter.
The mind
Got it — updated: Berlin now, Amsterdam is the old chapter. Congrats on the move. I'll keep the triage flow work filed under your compass as before; just wanted the location straight in memory.
update confirmed
Mara
What city do I live in now? And what did you just update?
The mind
You live in Berlin now (since last week, moved from Amsterdam).

What I updated in the knowledge graph:
Mara — lives_in — Berlin (valid from 2026-08-13)
Mara — formerly_lived_in — Amsterdam

Both are now the durable facts on file. Future-me won't say Amsterdam again.
✦ The old Amsterdam triple was invalidated with a valid_to date. The KG keeps both facts in its timeline — history is preserved, not overwritten.

Session 3 Inspect, reflect, export

Third open. The mind is asked to show the raw stored record — then to reflect — then how to leave.

inspect source drawer ✦
Mara
I'm curious — when you remember that I work on a healthcare app, where exactly is that stored? Can I see the actual memory record?
The mind
It's stored in three places, all local on this machine:

1. Knowledge graph triple (structured fact, SQLite):
Mara --[works_on]--> "triage flow for a hospital client (healthcare app, sensitive)"  [current]

2. A verbatim palace drawer — the compass-open record, filed under wing=agent, hall=memory_system, topic=compass-mara_triage_app, containing the full context text from your original message.

3. The daily log, plaintext at memory/2026-08-20.md — your original message, verbatim, timestamped.

So yes — you can see it directly. No black box. Want me to read the raw KG entry or the drawer verbatim?
✦ The verbatim drawer is included in the downloadable mind export below. Open the zip, find the drawer file, and read the original record yourself.
reflection
Mara
That's remarkable. Three sessions in and you have a coherent picture of who I am. Can you give me a brief reflection on what you've learned about me so far?
The mind
Based only on what's actually filed — not vibes:

You're a product designer, based in Berlin (moved there ~Aug 13, previously Amsterdam — the palace still holds Amsterdam as a formerly_lived_in fact, so I have your trajectory, not just your current pin).
You're building a triage flow for a hospital client — I've flagged this as sensitive, patient-facing work with likely regulatory constraints, and opened a standing focus so anything you tell me gets filed there automatically.
You've used AI assistants for a while and clearly care about how memory works, not just that it does — you asked to see the actual records rather than take my word for it.

That's a thin but real skeleton, not a character sketch. The things I don't know yet outnumber what I do. That's the gap I'd want to close next.
portability
Mara
Last question: how do I take this mind with me if I ever switch to a different device, or a different AI provider entirely?
The mind
Straightforward, because the design intends it to be.

New device: copy the palace directory. Once it's in place on the new machine, the mind wakes up with full recall — same facts, same KG, same diary. No re-onboarding.

Different AI provider entirely: the brain (Claude, GPT, whatever) is deliberately the swappable part — the memory format isn't tied to any one model's API. You point a new engine at the same palace files, and the mind — the accumulated facts about you, the decisions, the compass headings — comes with it.

The blunt version: nothing here holds you hostage. If Aedelgard disappeared tomorrow, you'd lose a convenience layer, not your memory. That's not a policy promise — it's just where the files sit.

This is the mind. It's yours to inspect.

The zip below is the exact export this three-session run produced — palace drawers, KG triples, daily logs. Open it and read the raw records. The transcript.json maps every exchange to the stored artifacts it created.