
Names, masks, keys, proof-of-personhood. The seventh flag was identity. This is the eighth: not who you are called, but what residue of work you allow to speak for you.

Self-custody for keys. This essay is self-custody for traces.
In 2019 I started planting flags: residency, banking, assets, a company. In 2022 I argued that names bestow capabilities only if meaning is attached in both the physical and the digital realm. In 2026 I called identity the seventh flag — own the registry, own the graph, own the agents, choose your proofs.
The missing flag was sitting in the same room the whole time.
Whoever holds the residue of your work holds a more useful you than whoever holds your legal name.
Not the passport. Not the ENS string. The dated notes, the agent sessions, the transcripts, the unsent drafts, the claims you revised in March and quietly dropped in June. That trail is how an institution, a model, or a future you will reconstruct intent. It is also how this site has always been made: a private store, and a public projection that is a fraction of it.
Thanks to Igor Schwarzmann — a friend — for the phrase memory as medium (April 2026). I take it from here through flag theory, a long-running personal store, and the fact that agents now write more of that store than I do.
The eighth flag
Flag theory diversifies jurisdictions so that no single state owns your life. The Network State made citizenship a subscription. Identity made the self diversifiable: names, graphs, faces, agents, proofs.
Memory does not diversify the same way. You can hold three passports. You cannot hold three pasts. You can project different views of one past. That is the whole game.
Eric Hughes, 1993: privacy is the power to selectively reveal oneself to the world. The warlocks in True Names feared that whoever learns your name gains power over you. In 2026 the cheaper attack is not the true name. It is the true log — ninety days of mail, calendar, and agent traces, indexed, and ready to autocomplete you.
Vendor “memory” is custodial banking of thought. ChatGPT Memory, Gemini Personal Intelligence, the helpful toggle that “remembers you”: those are correspondent banks. Convenient. Not yours. If you cannot export the substrate and the rules that govern what may leave it, you do not have a medium. You have an account.
I already made this argument for money. Be your own bank was about keys. This is about traces. The failure mode is identical: you discover you were a customer the week the correspondent freezes.
How I actually do this
I do not run a product. I run a store.
Dated research notes, people files, transcripts, clippings, a property schema so machines can filter before they read, and a hard rule: most of it never publishes. This blog is not the store. Capture the Flag is a rendering — selected claims, selected evidence, a date window, a human sign-off. The vault stays published: false unless I flip the bit.
That split is constitutional, not cosmetic. Composition still happens. What changed in 2025–26 is the volume of involuntary exhaust. Agents write session logs. Whisper writes transcripts. Clippers write captures. If I treated every artefact as a draft post, I would be working for the trail — green GitHub squares for thought.
So the job is no longer “what essay should I write for this audience?” It is: what derivative of work already done is appropriate here?
That is not laziness. It is the same discipline as selective revelation. A client briefing, a Ghost draft, a private digest for myself — three audiences, one substrate, three different legitimate lies of omission. Goffman was precise: the self in public is a stage. The honest claim is not “raw mind.” It is what I chose to make visible from work that actually happened.
Three jurisdictions, not three features
Think like a flag, not like a feature list.
| Jurisdiction | What lives there | Who may enter |
|---|---|---|
| Substrate | Raw traces: notes, sessions, mail, calendar, recordings | You, and tools under your key |
| Interpretation | Themes, open questions, claims, confidence, provenance | You, plus a model in an environment you trust |
| Projection | Posts, briefings, profiles, portals | A named audience, under written rules |
If you skip interpretation, you publish a dump. If you skip substrate ownership, interpretation happens on someone else’s disk. If you skip projection rules, you get vibe-marketing with a vector store.
Helen Nissenbaum called this Contextual Integrity: privacy is appropriate flow, not binary secrecy. A toggle that is on or off is not integrity. Integrity is: this class of fact may move to this audience, in this form, with this marker of who wrote it. Redaction should be acknowledged — a sentence that something was withheld — not silently erased. Silent erasure is how a “transparent” memory product becomes a lie.
Expose the shape of work, not unrestricted access to the mine. Shape is: what I am circling, what I dropped, what I now claim, how sure I am, over which months. The mine is the transcripts and the unsent drafts. Shape can travel. The mine should not.
Agents make this urgent
A context window is RAM. Memory is what survives the session. RAG is access to a corpus you may not own. None of those three is your memory.
What is new is that the agent is now a writer of substrate. In Harness, Loop, Graph I argued that the model is a rented brain and the harness is the jurisdiction. Durable state lives in files the next session can read. That is good engineering. It is also a pile of highly structured, highly embarrassing, highly useful text — unless you treat it like a wallet.
Conductance already drew the line for money: no raw PAN in long-term agent memory; research agents ≠ spend agents; kill switch in one runbook. Add a third identity:
| Agent | May |
|---|---|
| Research | Read substrate, draft interpretation, never pay, never post |
| Spend | Hold scoped credentials, never see the family calendar |
| Publication | Render a briefing from already-approved interpretation, never raw traces |
An agent that may draft should not be the agent that may pay, and should not be the agent that may hit Publish. Collapse those three and you have built a correspondent bank that also has your card and your blog.
Local, exportable files beat a vendor memory toggle. Markdown plus git is boring on purpose. Boring is how keys survived. Boring is how notes should survive.
Projection rules, or it is not a medium
Readers — including future you — need a rulebook, not a vibe:
- What classes of data never leave (health, family, credentials, unpublished research)
- Human-approved versus agent-drafted, marked as such
- Date window (this quarter, not “everything since 2014”)
- Audience (self / collaborator / client / public)
- Confidence, or a refusal to fake it
Without that, you are performing transparency. With it, you have a medium: the same store can feed this site, a closed briefing, and a private digest without lying about what was omitted.
Vanity metrics on the public face will corrupt the store. If the scoreboard rewards volume of snapshots, you will produce snapshots. I have watched that happen to “shipped” tickets that never shipped. The constitutional fix is the same as published: false by default. Projection is opt-in. Capture can be passive. Publication cannot.
I do not want a SaaS that is my memory. I want portable substrate, interpretation I can re-run, and surfaces I already own — a blog, a vault, a mail account. If the clever layer dies, the notes remain. That is the only product test that matters.
What I will not do
- Auto-publish agent exhaust. That is a leak with a product name.
- Sync the whole private store to a site, a CRM, or a “second brain in the cloud.” Dumping is not sovereignty.
- Treat a knowledge graph as a face. Graphs are data. People read claims and time.
- Put secrets, health, family, or payment credentials in long-term agent memory. Shape, not substrate.
- Confuse a vendor memory toggle with Contextual Integrity.
- Pretend a multi-month trajectory is a window onto the soul. It is a stage I built from real work. That is already a high bar. It is not omniscience.
Final thoughts
The identity essay ended on proof-of-personhood as the scarce asset of the AI age. Memory is the other half of that scarcity. Anyone can generate a polished artefact. Fewer people can show a trajectory of binding and revising claims without handing over the mine.
Own the registry. Own the keys. Own the traces. Choose the proofs — and the omissions.
Plant the flag on the residue. Keep the mine.
The website and the information contained therein are not intended to be a source of advice. Memory architectures have privacy, employment, and legal consequences; verify what you store, what you project, and under which jurisdiction.
Addendum: editorial assembly from a Markdown vault with AI assistance.


