Evolving understanding needs re-distillation, not composition

Type: note · Status: seedling · Tags: learning-theory, context-engineering

A note graph distributes knowledge across composable fragments — each note makes one claim, links provide traversal. This works for durable knowledge: the consumer picks what to load, coherence is local to each note.

When understanding is evolving and a consumer needs the whole picture — current beliefs, tried approaches, ongoing strategy — fragment reconciliation may not be feasible. Loading many notes, identifying which are current, and assembling a coherent view competes for context budget on two dimensions: volume (loading N fragments) and complexity (discriminating current from stale, resolving tensions). As fragment count grows, reconciliation can exceed effective context — making the operation infeasible, not merely expensive.

A pre-distilled narrative sidesteps this. The author — human or agent — performs reconciliation once and produces a single document sized for effective context. In distillation terms: accumulated understanding is the source, the consumer needing the current picture is the target, and the narrative is the distillate. When understanding changes, you re-distill — holistic rewrite — rather than let fragments diverge until reconciliation breaks.

Why composition is expensive here

At least three properties of evolving understanding push fragment-based reconciliation toward the effective context boundary:

  1. The whole picture must be loaded at once — before addressing their actual task, a consumer must load many notes, infer currency, and reconstruct connections. Each fragment competes for the same context budget as the task itself.

  2. Discrimination complexity grows with change rate — when new evidence contradicts earlier understanding, notes get added or revised individually. The consumer must then discriminate current from stale across the whole set. The faster understanding evolves, the harder this becomes: a note may be mostly current but wrong on one point corrected by a later note — a signal easy to miss.

  3. Coherence requires simultaneous loading — each note can be internally coherent while the set is collectively inconsistent. Resolving this requires holding all relevant notes in context at once and reconciling tensions between them. This is where the feasibility constraint bites hardest: reconciliation demands context proportional to the number and interdependence of fragments.

The general pattern

The pattern applies when: - Knowledge is accumulating from ongoing work, not static - A consumer needs the whole picture to act, not just a slice - Value is consumed (informs current action) rather than accumulated (builds a long-term graph)

Examples: onboarding documents tracking a changing system, investigation summaries during incident response, evolving design rationales during architecture exploration, theory documents during engineering campaigns.

The lifecycle is workshop, not library — the narrative lives and dies with the period of active evolution. When understanding stabilizes, insights should be extracted into durable notes, a second distillation into a different form for a different consumer.

Evidence: theorist

The theorist skill (MIT, blader, 2026) implements this pattern as THEORY.MD: a single narrative, holistically rewritten rather than appended, updated when understanding shifts rather than when code changes, capped at 200 lines to enforce concision. When earlier understanding is superseded, the document records the shift ("initially X, but Y revealed Z") rather than preserving the old version — it always reads as a present-tense narrative.

Theorist is illustrative rather than confirmatory: it demonstrates the pattern's mechanics but offers no controlled evidence that re-distillation outperforms composition. The strongest signal is indirect — teams using theory docs report easier mid-project onboarding, because there is one document to read instead of a trail of notes and chat history. Whether this holds at scale or for agent consumers remains open.

Open questions

  • Can agents perform holistic rewrite reliably, or does re-distillation require human judgment about what's still true?
  • What's the right extraction bridge when the evolution period ends? The narrative's insights need to become durable notes, but extraction is itself a distillation step that could lose context.
  • How does a size constraint interact with scope? Forced concision is valuable — it IS distillation — but complex situations may resist compression into a single document.
  • When does the rewrite cost exceed the reconciliation cost it prevents? If fragments are few and stable, a consumer can reconcile within effective context. The threshold depends on fragment count, change rate, and interdependence.

Relevant Notes: