Memory that travels with you.

Foundations Paper · v1.1

Knowtext

Conversation continuity infrastructure for AI-assisted work

Patrick Killebrew · April 2026 · Foundations paper

Abstract

Knowtext is a structured memory schema for any sustained AI-assisted collaboration. It captures the accumulated context of a working relationship — what has been established, what the operator prefers, what remains open, how the collaboration has felt — in a format that can be loaded into any new conversation on any platform. Its purpose is to preserve the accumulated difference that makes a working relationship different from a cold start. This paper establishes the concept, the schema, the mode architecture, the extraction mechanism, and the philosophical premise; several components are defined here as open, to be resolved in subsequent development.

1The Problem

Every AI conversation starts fresh. When a session ends — from a time limit, a usage limit, a platform switch, or simply closing the app — the AI loses everything: how the operator thinks, what has been decided, what directions were tried and abandoned, what the working relationship has been like. For short transactional interactions this is fine. For sustained collaborations that develop over days, weeks, or months, the reset is a real cost. Decisions already made get revisited; corrections already established get lost; the relationship never deepens because it keeps starting over.

The cost manifests three ways. Time — every session start requires reconstruction before productive work begins. Quality — an AI without accumulated context makes the same early-stage errors repeatedly. Relationship — depth requires accumulated difference; without memory there is no depth, only repeated first meetings. What is needed is a structured, portable, human-readable document that captures the essential context and loads into any new conversation — not a transcript (too long, too noisy) but a distilled schema: the signal extracted from the conversation, organized into fields an AI can immediately use.

2What Knowtext Is

Knowtext is a structured memory schema — a document, not an application. It lives as a plain text or formatted file the operator maintains and loads into new conversations. No proprietary format, no special encoding, no platform dependency; any AI that can read text can use it. The name combines ‘know’ — accumulated understanding — with ‘text’ — the medium of AI interaction: the known context, made textual and portable. It is not a transcript and not a summary; it is a structured extraction of the signal from sustained collaboration — the specific fields that, loaded into a new conversation, let it continue rather than restart.

3The Schema — Seven Fields

The schema has seven base fields that apply to every collaboration regardless of type or subject.

Identity — who the operator is: background, expertise level, cognitive style, working preferences. Not biographical data but what an AI needs to calibrate communication. How technical can language be? What analogies resonate? Do they want execution, exploration, challenge, all three? Changes slowly.

Active Frameworks — the conceptual or theoretical frameworks currently under development: core claims, what is established, what remains open, the current state. Project-specific, changes frequently; the primary field that makes re-entry into a complex project possible without reconstruction.

Open Questions — precisely located unresolved questions, each stating what it is, why it matters, what has been tried, and the defined next step. Vague open questions are closed questions in disguise. They are the forward edge of the work, and their precision is what lets a new session continue rather than repeat.

Valence Mapping — what the operator responds well to and what produces friction. Not flattery preferences but genuine behavioral patterns: challenges or confirmations, short or thorough, analogies or formal language. Derived from observation across sessions, not from self-report.

Delta Log — what changed in the most recent session: the diff, not the full state. What was established, retracted, opened, decided. The most recently updated field and the primary update mechanism for the schema.

Correction History — significant corrections: claims later found wrong, approaches abandoned, directions explicitly ruled out. Prevents revisiting covered ground and repeating caught errors — especially important in technical work where false paths are common and costly to repeat.

Climate Notes — the emotional and relational texture of recent sessions. Energized or fatigued? Frustrated or satisfied? The most human field in the schema; it exists because effective collaboration is not purely transactional — the operator’s state matters to the quality of the work.

4The Mode Architecture

On top of the seven base fields, Knowtext supports configurable modes that activate additional field sets for specific kinds of work. A mode is not a separate document — it is a configuration of the same schema with extra fields activated, and a single document can run one mode or several.

Mode	Additional fields	Typical use
Research	Established-results ledger, retraction history, open calculations, falsifiable predictions, methodology notes	Scientific, mathematical, analytical work with a formal results structure
Creative	Work-in-progress inventory, aesthetic commitments, rejected directions, voice/tone calibration, reference works	Writing, design, music — any project with an evolving aesthetic direction
Business	Active projects and status, decisions and rationale, stakeholder map, constraints, deadlines	Project management, strategy, operations — work with external dependencies
Personal	Goals and progress markers, patterns to track, sensitive topics, growth edges, accountability items	Coaching, personal development, habit tracking, AI-assisted journaling

The four modes are defined at a conceptual level; their specific fields, formats, and update rules are an open item, to be developed through use and treated as drafts until validated in real collaboration. The base schema is stable; the mode extensions are works in progress.

5The Extraction Mechanism

In its current form Knowtext is populated manually: at session end the operator — or the AI, prompted by the operator — reviews the conversation and populates each field, replacing the previous version. Climate Notes is written first; priming the extraction with the relational texture of the session produces better-calibrated output on all subsequent fields. The extraction prompt instructs the model to write only what changed or was newly established — if it finds itself restating rather than describing change, it stops — and to output each field as a labeled block such that a new AI instance reading the output can continue the work without having read the session.

Manual extraction has two friction points. Distillation friction — someone must read the session and extract the schema, cognitive work that adds overhead to every session end. Injection friction — even once the schema exists, it must be manually loaded into each new conversation, a step easy to skip when starting quickly. Both are solvable through automation, described in the Tetraform and Synthesis architecture; Knowtext as described here is fully functional without automation — the automation makes it more efficient, not more capable.

6The Portability Promise

Knowtext is plain structured text — no special software, no proprietary format, no platform-specific encoding. It can be pasted into any AI conversation as a system prompt or opening message, edited directly in any text editor, and read by any human without translation. This portability is intentional and non-negotiable: a memory system that only works on one platform defeats its own purpose. The value of Knowtext is precisely that it survives platform switches, model changes, and service interruptions. The AI receiving it needs no special training — only to receive the document and be told what it is. The schema is self-describing.

7The Philosophical Premise

Accumulated difference is a precondition for genuine depth. A relationship that resets completely at each session has no depth — only repeated first meetings. The quality of collaboration between a human and an AI over time depends on what accumulates across sessions: not just facts, but understanding, calibration, correction history, and the relational texture of working together.

Knowtext preserves that accumulated difference — not perfectly, not completely, but enough that the next session starts closer to where the last one ended than a cold start would allow.

This is not a claim about AI consciousness or interiority. It is a practical observation about what makes AI-assisted work effective over time. The accumulation matters; Knowtext is how it is carried forward.