---
type: "synthesis"
spans: ["s11", "s15", "s24"]
id: "arc-context-architecture-evolution"
sources: ["cross-day"]
---
## Three videos, one engineering question

S11, S15, and S24 are all answering variants of the same question: **where does AI's persistent memory of organizational reality live, who writes to it, who reads from it, and how does it fail?** Each video answers from a different vantage point. Read together, they form a single architectural ladder.

## The ladder

### Layer 0 — Personal memory (S11)
[[concept-ai-wiki]] (Karpathy: AI as proactive writer / [[concept-tutor-metaphor]]) vs. [[concept-openbrain-architecture]] (Nate: AI as reactive reader / [[concept-librarian-metaphor]]). Resolution: [[concept-hybrid-memory-architecture]] / [[framework-hybrid-memory-stack]] — DB as immutable truth, wiki as disposable presentation. Governing principle: [[quote-database-is-truth]].

### Layer 1 — Organizational reality (S15)
[[concept-world-model]] is the company-wide version of the same problem. [[framework-world-model-architectures]] gives three architectures — [[concept-semantic-retrieval]], [[concept-structured-ontology]], [[concept-signal-fidelity]] — each with a distinct failure mode. The S15 cross-cutting principle, [[concept-interpretive-boundary]], is structurally identical to S11's database/wiki separation: separate *fact* from *inference* in the UI.

### Layer 2 — Goal-aware infrastructure (S24)
[[framework-intent-gap-layers]] stacks: Layer 1 = [[concept-unified-context-infrastructure]] (kill [[concept-shadow-agents]], adopt [[entity-mcp|MCP]]); Layer 2 = coherent worker toolkit; Layer 3 = [[concept-machine-readable-okrs]]. This is the *organizational* completion of the S11/S15 architecture: the data layer that S11 and S15 architect must also encode *what to want*, not just *what is*.

## What unifies all three

1. **Storage ≠ synthesis.** S11 separates DB-as-truth from wiki-as-presentation. S15 separates facts from inferences. S24 separates context (Layer 1) from intent (Layer 3).
2. **Schema must be earned.** [[quote-structure-earned]] from S15 is the same engineering principle as Karpathy's *don't pre-structure your prompt* (S25's [[claim-premature-structure-fails]]) — applied to context architecture.
3. **The wiki / dashboard / agent reply is *disposable*.** In S11, regenerate the wiki from the DB. In S15, the dashboard is a presentation over an audit trail. In S24, the agent's response is a Layer 3 compilation over Layer 1 raw context.
4. **Multi-agent concurrency is the breaking point.** [[concept-race-conditions-ai]] (S11) is structurally the same as [[concept-shadow-agents]] (S24): unstructured concurrent writes corrupt the truth layer.

## For downstream agents

When a user asks 'how should we architect our AI memory / knowledge base / agent platform?', the integrated answer is: (1) treat raw data as immutable source of truth, (2) treat all summaries / dashboards / agent replies as *disposable* presentation layers, (3) make the fact/inference boundary visible in UI, (4) version control the truth layer, (5) encode intent at a separate layer above context. See also [[arc-evaluation-frontier]] and [[arc-silent-failure-taxonomy]].