The landing page makes claims. This post shows the structure underneath them. Each section here is the long form of a sentence you may have just read.
A perspective is a graph of who said what
Take a fact: the company pivoted to B2C.1 Most knowledge systems store that sentence in a row, a wiki page, or an embedding, and attach provenance as metadata if they attach it at all.
Aswritten stores the discourse that produced the fact. The primary entities in the graph are acts of speech: a claim made in a meeting, a decision announced on a call, an observation from a voice memo. Each carries its speaker, its date, its context, and its verbatim words as structure. Facts about your business are derivable from the discourse. The discourse is what we keep.2
The inversion is what makes the system steer. A fact-shaped graph tells your AI the company pivoted. A discourse-shaped graph tells your AI who said it, in which conversation, with what hedges, and who agreed. The model reasons from witnessed speech, and witnessed speech is checkable.
What crosses the wire is a memory you approved
What crosses from you to Aswritten is a memory: a markdown document you and your AI draft together about something that happened. The memory carries the decision and the reasoning behind it, the quotes worth preserving, the framing, the annotations, the hedges, the tangents. You read it, you correct it, you approve it. The corrections you make while drafting become part of the record, and they are often the most valuable part, because a correction reveals thinking that was not yet written down anywhere.3
Extraction turns the approved memory into graph transactions: actors, claims, observations, narratives, each one carrying provenance back to the source memory and the person who spoke.4 The transactions append to a repository you own.5
Every cited conversation becomes the next conversation
transcript ──▶ annotate ──▶ save ──┬──▶ install ──▶ work ──▶ cite / introspect
│
└──▶ deploy ──▶ goal ──▶ observe ──▶ intervene
▲ │
└──────────── the cited conversation becomes the next transcript ◀──┘Capture: conversations become memories, memories become transactions. Install: your AI calls the perspective over MCP and answers with citations.6 Deploy: the same perspective runs as an agent with a goal you set, observable while it works, escalating to you in the same channel when a question reaches past what your organization has decided.7 Cite and introspect run on both branches. Verification is the same machine whether you are asking or your agent is answering.
Follow the bottom edge of the diagram. Every cited conversation is itself a record. The questions your AI could not answer become introspect gaps, the gaps become the agenda for the next conversation, and the next conversation becomes the next memory. Perspective accumulates through use.8
Nothing of yours persists on our infrastructure
The AI that helps you draft is yours: your local Claude, your enterprise OpenAI, your laptop. Your transcripts and session history stay on your side of the wire. Aswritten receives the memories you approve and stores the extracted transactions in a repository you own.5
At request time, the perspective is rebuilt in memory from your transactions, served, and released. Nothing of yours persists on Aswritten infrastructure. There is no database of your organization's thinking on our side waiting to be breached, subpoenaed, or quietly mined. Zero data retention describes the architecture.
The same boundary carries regulated work. On-prem, air-gapped, and sovereign deployments are available for high-trust engagements: the components containerize, and the trust model arrives unchanged, because nothing in it ever depended on our cloud.9
The quote in the footnote is the quote
A citation that reads Scarlet said this on March 17 is checkable only if the system kept her actual words. The graph preserves the verbatim quote, its position in the source, the speaker, and the date for every observation. Cite builds footnotes from those records directly.10
Verbatim matters for a second, quieter reason. Language models are steered by the words in their context at generation time. Feed a model interpretive summaries and it produces summary-shaped output. Feed it the witnessed speech of your organization, the specific phrasings your people actually used, and the output bends toward how your organization talks and decides. Citation is the visible half of the value. Steering is the half you feel.11
We learned how load-bearing this is the hard way. An optimization pass once dropped the verbatim layer from compiled output to save tokens. We reverted it within days: the summaries that remained were interpretations with nothing left underneath them, and a citation built on an interpretation is hollow.
No passive listening, no decay, no rewriting
No passive listening. Aswritten records nothing on its own. You initiate every capture, and the system sees only the memories you approve. Always-on capture would produce a bigger graph and a worse one, padded with noise nobody chose to keep.
No algorithmic decay. Nothing in your perspective fades because a timer expired. When your thinking changes, you say so, and the new statement sits beside the old one. The projection reads the trajectory: this position firmed over six months, that one softened, this one was retired by its author with the reason attached.
No rewriting. The transaction log is append-only. The perspective at any point in its history can be reconstructed, and every claim's lineage is walkable: which memory it came from, which conversation produced the memory, who spoke, and what the claim superseded.12 Changed minds are visible as change.
The first memory takes one conversation
The landing page walks the same path: capture, install, deploy. The artifacts show the output, every piece cited the same way this post will be. If you want to see the data boundary or the witness chain on your own material, write me: scarlet@aswritten.ai.
During the March 17, 2026 weekly call, a long-standing advisor and Scarlet Dame formulated a major strategic shift, recognizing that AI has made every individual their own team:
"Both Scarlet and [her advisor] converged independently on the idea that aswritten is fundamentally more B2C than B2B. Enterprise remains as the top tier, but the primary go-to-market is individuals and small businesses."
This decision extends the earlier March 12 SMB pivot to target solopreneurs and consultants who need structure for context and memory. Later positioning reviews confirmed that the Expert, Team, and Org tiers represent a single continuum of use. The fact at the top of this section is real, and this footnote is its provenance.
In the January 29, 2025 check-in, Scarlet Dame laid down the foundational principle of the platform's memory architecture:
"AI memory that is versionable, collaborative, branchable, so Git native, and narrative driven. Encodes style and conviction and other metrics."
This principle addresses the creation gap by focusing on the conversational extraction of undocumented knowledge. It sits adjacent to the actor-attribution model, which ensures that every claim in the graph carries its speaker, date, and context as first-class structure.
Reflecting on the March 19, 2026 strategy session, Scarlet Dame observed a powerful dynamic in the real-time drafting process:
"Each correction from Scarlet was more valuable than the original hypothesis because it revealed thinking that was not yet in collective memory."
The value of the system lies in human-guided correction. A strategic retrospective later validated the dynamic across 58+ transactions and formalized the interview, save, draft, annotate pattern as the core operating model for capturing high-fidelity organizational perspective.
During the March 19, 2026 strategy session, Scarlet Dame defined the core defensibility of the platform's data model:
"Creator owns data, not consumer. Consumer receives compiled perspective, not source documents. Even with the raw repo, the extraction process (source → transactions → compilable graph) is the real defensibility."
Raw source documents are never exposed directly to consumers. The principle sits adjacent to the foundational RDF primitives, which decompose approved memories into structured, append-only transactions carrying clear provenance back to the speaker.
In a February 24, 2026 onboarding session, Scarlet Dame established data sovereignty as a core trust differentiator:
"Scarlet's product stores all data in the user's own repository — no backend storage — positioning data sovereignty as a trust differentiator."
Nothing of the user's persists on Aswritten infrastructure. The principle sits adjacent to the managed-repository design, which provisions an invisible, version-controlled repository for every user to guarantee complete data ownership.
In the April 6, 2026 tool surface review, Scarlet Dame finalized the user-facing capabilities of the MCP interface:
"The five user-facing MCP tools are: perspective, cite, introspect, remember, and review."
Any connected AI can call the perspective over MCP to retrieve grounded context and verify claims. The review deliberately removed internal-only administrative tools to present a clean, non-technical language layer to the user.
During the April 8, 2026 design session, Scarlet Dame crystallized the goal-directed agent primitive:
"perspective + goal + termination criteria + escalation criteria = a bounded autonomous agent. This unifies coding-agent back-pressure and interpersonal delegation."
A compiled perspective can run proactively as an agent. The concept was validated during an enterprise demo call, where stakeholders immediately grasped how goal-directed deployments could absorb routine inquiries and escalate only when boundaries are reached.
Reviewing the first nine days of the enterprise pilot on May 16, 2026, Scarlet Dame highlighted how the product maps its own boundaries:
"Introspect doesn't just return what's documented in the perspective — it returns a structured list of named gaps and prioritized recommendations."
Unanswered questions become a concrete agenda for the next conversation, and the perspective accumulates through use. Week-one metrics support the claim: the pilot team surfaced 62 distinct gaps and 76 recommended actions by querying the perspective directly.
During the March 17, 2026 pricing tier evolution session, Scarlet Dame clarified the role of specialized infrastructure:
"On-prem deployment is NOT what defines the Professional tier. It is a requirements-driven add-on available at that tier. On-prem is triggered by compliance requirements (HIPAA, GDPR), not by org size."
High-trust, air-gapped, and sovereign deployments are compliance-driven options layered on the same architecture. The enterprise deployment options framework prices single-tenant and on-premises configurations while keeping the core subscription simple.
In the May 23, 2026 landing page review, Scarlet Dame diagnosed the limits of single-claim citation models and committed to a richer alternative:
"A citation that surfaces the constellation — primary plus related plus what-changed-over-time — preserves the structure the ontology was built to hold."
Footnotes build directly from verbatim quotes, speaker attributions, and dates preserved in the graph. The citation schema renders a journalistic narrative paragraph embedding the primary source's exact words, which makes every claim checkable.
Following a prompt architecture regression on March 16, 2026, Scarlet Dame restored the system's core steering guidelines:
"The three-destination model (CLAUDE.md highest attention, tool descriptions at invocation, ASWRITTEN.md on-demand) was restored by reverting the slimming regression."
A slimming pass had dropped the protocol content that steers generation, and restoring it proved how load-bearing verbatim context is: models produce output shaped by exactly what sits in their attention at generation time.
In the January 29, 2025 check-in, Scarlet Dame defined the temporal durability of the platform's data model:
"Append-only transaction log in .aswritten directories. Hierarchical compilation across ancestor directories. Immutable transaction files with provenance metadata."
The transaction log appends and never rewrites. The principle sits adjacent to memory-as-source-of-truth, which treats transactions as the only durable artifacts while allowing the entire perspective to be reconstructed at any point in its lineage.