You've been working on this for weeks across a dozen sessions. The thinking is preserved in conversation history. But nothing marks which moments were turning points. The archive is flat.
You open an old thread and scroll. Somewhere in there you made a decision that changed everything. But it looks exactly like the fifty messages around it. You can't find it. Or worse, you find it and realize you contradicted it three sessions later without noticing because neither you nor the AI remembered it was important. The thinking happened. The importance wasn't recorded.
AI treats every message with equal weight. A casual aside and a foundational decision occupy the same position in context. When the context window fills or a new session starts, the AI has no way to know which pieces to preserve and which to let go. This is semantic flattening: the loss of importance hierarchy across time.
Semantic Hierarchy fixes this by teaching you to mark what matters. You learn to identify the moments that govern future work (decisions, constraints, realizations) and record them in a format the AI can use to maintain continuity. The result is context that survives across sessions because you told it what to keep.
A working semantic hierarchy system for your own practice. You'll take a real multi-session project, identify the turning points you missed, build a protocol for marking decisions and constraints as they happen, and test it across at least three sessions. The deliverable is a repeatable practice: you know what to mark, when to mark it, and how to format it so the AI carries it forward.
Deliverable: A semantic hierarchy protocol tested across multiple sessions on a real project, with evidence of context preservation.
Full scaffolding: format and trigger criteria provided
You open a session from two weeks ago. Somewhere in the middle you made a decision that restructured the entire project. It looks exactly like the hundred messages around it. You scroll, skim, give up. The decision happened. The record doesn't know it was important. This is semantic flattening: every message enters at the same weight and stays at the same weight. The sharp turns are buried in the filler. In September 2024, one practitioner requested "the full scope of our ideation in a copy-able format as a savepoint before we get lost" in six consecutive sessions before the protocol existed. The problem was clear before the solution was.
You learn Savepoint Syntax v3.2: the self-closing tag format, what each field does (category, function, timestamp, project, keywords, content line), and why a one-line crystallized thought beats a paragraph summary. You drop your first savepoints using provided trigger criteria: reframes ("wait, it's actually..."), structural decisions (naming, hierarchy, what goes where), connections between previously unconnected things, energy shifts when something clicks, drift detection when work has deviated from intent. The criteria are explicit. Your job is recognition, not invention.
Deliverable: Five savepoints dropped across provided session material, each with correct syntax and a trigger-criteria justification for why that moment qualified.
Reduced scaffolding: trigger criteria become guidelines
Marking turning points is a bilateral skill. The AI can recognize the shape of a potential turning point: a topic shift, a move from exploratory to declarative language, a new term introduced. What it cannot determine is whether that shift is a governing decision or a passing tangent. That determination requires knowing what you are trying to do. The AI compensates for your limited bandwidth by watching for turns. You compensate for its inability to judge weight by being the source of intent.
You practice on your own project in live sessions. You learn to recognize when the AI should propose a savepoint (it spotted a pattern you missed) and when only you can judge the weight (the shift only matters because of a constraint you declared three sessions ago). You work with intent context: project governance docs (CLAUDE.md, institutional memory files) that tell the AI what "important" means for this specific project. The same sentence is filler in one project and a governing decision in another. The trigger criteria from Week 1 become guidelines, not rules. Your judgment starts replacing the scaffolding.
Deliverable: A live session on your own project with savepoints dropped in real time, plus a reflection identifying which savepoints were AI-proposed vs. human-initiated and why.
Independence: you mark, traverse, and compile without guidance
Savepoints are not a journaling practice. They are the index to an archive. The full loop: mark turning points during sessions, persist them in conversation exports, traverse the archive later using savepoints as high-weight markers, compile grounded content from what you find. A savepoint dropped in January tells the traversal system in March where the thinking was, not just what was discussed. Without that marker, the connection between a naming decision and a copy rewrite three months later is gone.
You build the complete loop on your own project. You practice knowledge traversal: chronological reading of your own conversation archive with savepoints as waypoints. You learn what the savepoints enable downstream (traceability, context reconstruction, grounded compilation) and what they cost when they are missing (flattened archives, contradicted decisions, lost continuity). You produce one compiled document from traversal, a piece of real output whose content traces back through savepoints to the sessions where the thinking happened.
Deliverable: A working savepoint practice with an archive that has hierarchy, plus one compiled document produced from traversal of your own marked sessions.
Same conversation archive. One flat. One marked.
A 306MB JSON export of conversation history. Eleven projects in six weeks. Every one required the practitioner to stop working and manually request a "savepoint" before the thinking disappeared. The same corpus, two approaches to preserving what mattered.
"Please give me the full scope of our ideation in a copy-able format as a savepoint before we get lost in conversation."
The same request, six sessions in a row. "I'd like to generate a very detailed overview... that i can copy and save as a 'save point.'" Each one a manual intervention. The work of thinking interrupted by the work of preserving the thinking.
<Savepoint protocol_version:3.2 category:decision function:declaration timestamp:2026-03-12T04:32:47Z project:petersalvato.com keywords:taxonomy, structure, naming context:When state changes frequently, the structure itself should recurse rather than being captured at a point in time. # The three-tier taxonomy is locked. />
This course uses scaffolded independence. Knowing what matters is a judgment skill, not a mechanical one. You can't follow a checklist for it. Early in the course, the scaffolding is heavy: you work with pre-marked examples, identify what was marked and why, and practice marking decisions in guided exercises. The scaffolding thins as you go. By Week 3, you're marking your own real work with no guidance. The scaffolded model works here because the skill is developmental. You're building judgment, not learning a procedure. The supports are there until you don't need them, then they're gone.
This course applies the research published in the Semantic Flattening whitepaper. Why AI memory loses what mattered, and what human-marked importance fixes.
This course is in development. $499 per course. Payment plans available on all courses. Foundations (Course 01) is the prerequisite. See the full curriculum.
$499. Requires Foundations. Payment plans available.