Memory as Medium

What is it?

The thesis that memory itself — not essays, not posts, not notes, not messages — becomes a publishing medium, a professional identity surface, and a coordination layer for AI-mediated knowledge work. This is not a claim about better storage or smarter retrieval. It is a claim about a new kind of public object: a structured, bounded derivative of ongoing cognitive work that makes thinking legible without exposing the substrate it came from. It is also a claim about identity: how we remember shapes who we are, and how we publish memory shapes how others perceive us. Controlling the projection is an act of narrative self-determination. (What “memory” means here is deliberately unsettled — it is neither human recall nor machine storage but the structured residue of their collaboration. The agentic memory landscape maps the machine side of this question; an earlier essay explored the human side.)

Memory = activity + reasoning about that activity + time. Not a log, not a snapshot, but the structured trajectory of how understanding evolves.

The insight emerges from a convergence of three independent developments:

1. LLMs as boundary adapters generate structured cognitive exhaust as a byproduct of work and can translate one rich internal object into whatever external format a situation requires — email, brief, deck, post, Slack update. This is what makes cognitively richer internal primitives viable without the interoperability friction that killed earlier attempts like Google Wave.
2. User-owned data protocols like the AT Protocol have made it architecturally viable to store these records in personal repositories and render them through multiple views (AppViews) for different audiences. Once you have something worth storing, the question becomes: who owns it?
3. Agentic memory systems have matured to the point where agents can accumulate, structure, and reason about experience across sessions — memory is no longer just “what was said” but “what can be concluded.” This is the capability that makes the published object worth reading.

The result is a new design space: systems where private working memory is continuously accumulated from fragmented activity (notes, chats, docs, email, agent sessions, meetings), structured into interpretable objects (themes, questions, tensions, claims, provenance), and selectively projected into public or shared surfaces. The published artifact is not a finished statement but a bounded epistemic object — evidence of thought, not the substrate of thought.

The architecture

Memory as Medium requires three layers, each with distinct visibility and trust properties:

Private substrate. Raw activity stays local: agent traces, Slack messages, email threads, vault notes, meeting transcripts, calendar signals. This layer accumulates passively without requiring workflow replacement. The user does not need to adopt a new tool on day one — the system observes existing tools and derives structure from their exhaust. The design choice toward passive capture is deliberate: research on the photo-taking impairment effect demonstrates that actively documenting an experience reduces memory of it — the act of capture offloads the responsibility of remembering onto the device. Involuntary exhaust avoids this trade-off. The substrate accumulates without demanding attention, preserving the cognitive engagement it records.

Interpretation layer. A synthesis engine (LLM-powered, running locally or in a trusted environment) continuously builds higher-order memory objects from the private substrate: recurring themes, open questions, emerging claims, unresolved tensions, referenced artifacts, and confidence levels. This is where raw data becomes structured meaning. The interpretation layer is what distinguishes Memory as Medium from simple capture tools like digital gardens or timeline apps — the system does the stitching, not the reader.

Publication layer. Selected derivatives are projected outward as bounded, structured objects. Different audiences see different views of the same canonical record: a private reflective workspace, a collaborator briefing surface, a client transparency portal, a public professional profile, a periodic digest. The projection is governed by explicit rules grounded in Contextual Integrity — publish structured explanation of context, not the source context itself. What gets withheld is acknowledged honestly (redaction summaries), not hidden.

This architecture inverts the normal relationship between private and public. Most publishing systems ask: “what content do I want to create for an audience?” Memory as Medium asks: “what derivative of my ongoing work is appropriate for this audience to see?” The author’s job shifts from writing to curating projection rules.

Each layer of the architecture corresponds to a different primary commitment in the Agentic Memory design space. No single memory system covers all three — which is why Memory as Medium is an integration thesis, not a storage thesis.

The primitive

The canonical object behind any published artifact is not a post, not a document, not a note. It is closer to a hyperedge — a single object that simultaneously binds multiple themes, questions, tensions, claims, and evidence into a situated entanglement at a specific moment. The primitive is the relation, not the thing.

What we currently call publishing primitives — posts, documents, briefs, updates — are more accurately understood as derivatives: renderings of this canonical object for specific audiences and contexts. A blog post, a client update, a Slack summary — each is a projection of the same underlying memory snapshot, translated into whatever format the receiving context expects. The memory snapshot is the actual primitive. The formats we know are its projections.

When someone subscribes to a memory feed, they are not reading the author’s thoughts. They are watching how things bind and unbind in the author’s cognitive field over time — which themes cluster together, which tensions resolve, which questions persist, which claims strengthen or weaken. The graph’s evolution is the content.

This reframe is critical for trust design. The system doesn’t expose what you think; it exposes the shape of your thinking. That distinction is what makes legibility possible without surveillance. It’s the difference between giving someone access to your files and giving them a living view of your work state.

Concretely, this object — call it a memory snapshot — has two layers: a relation core (machine-tractable references to themes, questions, claims, tensions, artifacts, source contexts) and a narrative frame (human-readable title, summary, confidence, provenance, redaction summary, time window). The text summary is one rendering of the object, not the object itself.

The trust problem

The hardest challenge is not infrastructure. It is editorial legitimacy: why should anyone trust that a memory snapshot is faithful, not flattering, manipulative, or over-compressed?

The projection rulebook — the explicit logic governing how private substrate becomes public artifact — must be transparent. Readers should know: what kinds of data were excluded, whether a human approved the snapshot, whether it was agent-generated or human-authored, how much abstraction occurred, and what confidence the system has in each item. Without this, memory publishing risks becoming vibe-marketing. With it, it becomes a genuinely new medium.

The privacy architecture must be grounded in something more principled than a simple public/private toggle. Contextual Integrity (Nissenbaum) provides the right framework: privacy is about appropriate information flow, not secrecy. The projection rulebook encodes context-specific norms about what can flow to which audience and in what form. A client gets one view. A collaborator gets another. The public gets a third. Each is a legitimate projection of the same underlying work state, governed by rules that can be inspected.

The medium as stage

Memory as Medium creates a new stage, not a window into a backstage. Goffman’s dramaturgical framework is instructive: every social context involves performance, and publishing through any medium is inherently theatrical. The private substrate captures involuntary cognitive exhaust — but the publication layer is curated through projection rules. The author chooses what to project, how to frame it, what to redact. McLuhan’s insight applies directly: the medium shapes the message. A memory feed would inevitably develop its own performative conventions, just as every medium before it has. People would start working for the trail, not just working — the way GitHub’s contribution graph incentivizes commits for the sake of the streak.

What makes Memory as Medium a different kind of stage, not a transparent one, is the dramaturgical cost structure. The performance cost is lower (exhaust vs. composition), the signal is harder to fake (behavioral patterns accumulated over time vs. single polished artifacts), and the audience reads shape rather than content. The projection rulebook is the script — but it operates over material that was generated involuntarily, which constrains how much the performance can diverge from the underlying reality. The honest claim is not “this shows you how someone thinks” but “this shows you what someone chose to make visible from the residue of how they actually worked.”

Different metaphors for this system reveal a spectrum of curatorial intensity:

In practice, Memory as Medium would likely operate in the middle of this spectrum — somewhere between studio visit and exhibition. The audience experiences a curated arrangement, not raw access, but the material on display is derived from actual cognitive work, not fabricated for the occasion. Curated projection of involuntary material.

This is where Haltung enters — the German word that simultaneously means posture and conviction, taste expressed as stance. The projection rulebook is Haltung encoded as infrastructure: choosing which cognitive traces to make visible requires the same compressed cultural intelligence that Virgil Abloh applied when deciding which 3% of an existing design to change. Deep knowledge of the field, sensitivity to context, and the conviction to commit to an editorial position. Without Haltung, the system produces data dumps. With it, the curation itself becomes the signal — and curation, not creation, is the defining skill of AI-mediated knowledge work.

The social object argument

Memory as Medium is not just a productivity thesis. It is a media economics thesis. Every major social platform emerged at the moment when the cost of producing a shareable unit crossed below a threshold. Photos: film + darkroom + scanning → tap on phone (Flickr, Instagram). Short-form thoughts: blog post with title, thesis, hosting → 140 characters (Twitter). The desire to share was always there. The cost was the barrier.

Reasoning is next. Right now, making your thinking process visible requires enormous discipline — maintained zettelkastens, public working notes, detailed journals. The people who do it (Andy Matuschak, Maggie Appleton) get attention that is substantially discipline admiration. AI-augmented workflows collapse this cost toward zero. When you work with AI tools, reasoning traces are generated as a byproduct — session logs, knowledge graph updates, meeting syntheses, evolving positions. You don’t journal your thinking in addition to working. The journal IS the exhaust of the work.

When the cost drops, the attention shifts. Pre-Instagram, a well-composed photo signaled craft mastery. Post-Instagram, the craft signal weakened and the perspective signal emerged — what you notice, where you point your camera, how you see. The same shift applies: when everyone can produce visible reasoning trails without extra effort, the discipline premium disappears and the perspective premium emerges. What gets rewarded is what your trail reveals about how you think.

Jyri Engstrom’s social objects thesis provides the analytical framework: people don’t socialize then find objects — they gather around objects. Decomposing what makes a photo work as a social object reveals six structural components: surface (instant legibility), opacity of craft (skill visible in the result), projection space (viewer completes the meaning), completeness (works in isolation), engagement gradient (from like to deep conversation), and identity leakage (involuntary self-portrait). Reasoning traces score well on craft visibility, engagement gradient, and identity leakage — but poorly on surface legibility and projection space.

The design implication: layer, don’t replace. Don’t try to make reasoning a standalone social format. Keep existing social objects (links, documents, shared artifacts) as the surface. Add the reasoning trail as a structured depth layer — the why behind the what. Not a caption (performative), but structured residue (involuntary) — the EXIF data of the thinking process. The curation of which trails to make visible is itself an identity signal — and that curation layer is the product.

Why now

Google Wave attempted a version of this in 2009: a shared object that was simultaneously conversation, document, live process, and history. It was right about the primitive but wrong about the timing. Wave dissolved familiar categories (email, docs, chat) into a compound object and asked the entire ecosystem to adopt it. The friction at every boundary — interacting with people who didn’t use Wave — killed it.

Three things have changed. First, the supply-side cost of sharing reasoning has collapsed. AI-mediated work produces structured cognitive exhaust as a byproduct — the barrier that kept reasoning private is dissolving the same way the photo barrier dissolved.

Second, LLMs serve as boundary adapters: they can translate one rich internal memory object into whatever external format the receiving environment expects (email, Slack update, brief, post, API payload). This means the internal primitive can be cognitively richer than anything Wave attempted, because the system no longer requires the rest of the world to meet it halfway. Complexity becomes affordable — the More Work for Mother principle applied to data models.

Third, protocol infrastructure (specifically atproto’s repo + AppView + Lexicon architecture) makes it technically viable to store canonical user-owned records and render multiple views without duplicating data or rebuilding the social graph from scratch. As Bain Capital Crypto argued: “Many types of data have a lifecycle which begins as private but moves to public.” That lifecycle is the product. A solo consultant with three agents is already a multi-actor system — there is no phase transition between “individual” and “team.” The protocol doesn’t care about the composition; it cares about sovereignty and views.

Existing experiments

This thesis is not purely theoretical. Three related experiments already test parts of it:

Shev and the Observer project. Shev is a personal observer agent that extracts reasoning moves from Claude Code sessions (761 sessions, 150K turns, ~10M tokens) and meeting transcripts. It builds a cognitive terrain — a navigable map of what has been thought about, how intensely, and through what reasoning modes. Shev is the upstream capture layer that would feed a Memory as Medium system. The 12-type reasoning move taxonomy (validated across 10 transcripts, all types firing) provides the interpretive vocabulary.

Atmosphere. The visual layer of the Observer project — a WebGL nebula at igorschwarzmann.com/atmosphere/ that represents the current state of cognitive engagement as a spatial, temporal, ambient display. Atmosphere is what a public-facing memory surface might look like: not a feed of posts but a living visual representation of active work. It proves that cognitive state can be rendered as something other than text.

The vault itself. The Brain Dead vault — with its Obsidian notes, Claude Code sessions, Granola transcripts, Honcho memory, ccvault session archive, qmd search, and agent ecosystem (Shev, Sam, Carl, CJ) — is already a bespoke implementation of the private substrate layer. It accumulates cognitive exhaust from multiple tools, structures it through frontmatter schemas and wikilinks, and supports multiple views (Dataview queries, MOCs, published concept notes). The product question is whether this bespoke infrastructure can be generalized into something installable.

The business thesis

Memory as Medium implies a three-layer economic model that goes beyond productivity SaaS:

1. SaaS layer. Individuals and teams pay for private capture, synthesis, and memory. The wedge value is continuity and recall — fewer lost insights, faster synthesis before client moments, a durable trail of how concepts evolved.

2. Visibility layer. Clients, collaborators, and stakeholders consume AppViews over the user’s work state. This converts a permissions problem into a publishing problem: the client doesn’t need access to the files; they need a structured, continuously updated view of the work as it evolves. McChrystal’s shared consciousness framing is precise here: “not everybody can know everything, but you have to have enough linkage so that it pulls you into a similar relationship that the small team enjoys.” More data does not equal shared consciousness — linkage infrastructure does. AppViews are that linkage. Revenue could attach to per-workspace or per-collaborator pricing.

3. Network layer. As more proof-of-thought artifacts become public, discoverable, and subscribable, the system develops network effects. Professional identity is built from demonstrated cognitive engagement over time, not from curated status updates. The unit of social exchange is not “content” but “legible work state.” This is where the product stops being a tool and starts being a new graph of professional trust and visibility.

The wedge is practical: solve fragmented synthesis for AI-heavy synthesis workers and small teams. The category is ambitious: define a new publishable object for knowledge work, with social and professional surfaces emerging from that object.