---
type: concept
topic: agent-wiki
status: active
created: 2026-04-06
updated: 2026-04-07
source_notes:
  - "[[research/agent-wiki/sources/karpathy-llm-wiki]]"
  - "[[research/agent-wiki/sources/anthropic-building-effective-ai-agents]]"
  - "[[research/agent-wiki/sources/dissenting-views-on-persistent-agent-wikis]]"
derived_from:
  - workspace-initialize
confidence: medium
shared_concept:
---

# Compiled Knowledge Layer

## Definition

A compiled knowledge layer is the maintained markdown synthesis that sits between immutable source notes and future questions. Instead of rediscovering meaning from raw inputs on every query, the LLM incrementally updates durable pages that already contain cross-links, tensions, and prior synthesis.

## Why It Matters Here

This is the central architectural idea of the workspace. If it holds up, the vault stops behaving like a pile of notes plus ad hoc search and starts behaving more like a living, inspectable knowledge system.

## Supporting Sources

- [[research/agent-wiki/sources/karpathy-llm-wiki]] frames the wiki itself as the persistent artifact.
- [[research/agent-wiki/sources/anthropic-building-effective-ai-agents]] adds the constraint that this layer should stay legible and not become needless machinery.
- [[research/agent-wiki/sources/dissenting-views-on-persistent-agent-wikis]] adds the crucial warning that a compiled layer can become an error persistence mechanism if provenance and retrieval backstops weaken.

## Tensions

- A richer compiled layer increases reuse, but it also raises maintenance cost and ontology drift risk.
- The system is strongest when the wiki compounds useful synthesis, but weakest when it captures too much low-value residue.
- If the layer starts behaving like silent authority rather than visible interpretation, stale or poisoned claims can harden into infrastructure.

## Implications

- `index.md`, `log.md`, and concept pages are not admin overhead; they are part of the memory architecture.
- Outputs should be saved selectively when they add reusable understanding.
- Promotion to shared concepts should be sparse and earned.
- Retrieval against canonical notes should remain available as the verification and recency backstop.

## Related Concepts

- [[research/agent-wiki/concepts/governed-interpretation-layer|Governed Interpretation Layer]]
- [[research/agent-wiki/concepts/simplicity-before-autonomy|Simplicity Before Autonomy]]
- [[research/agent-wiki/concepts/obsidian-as-agent-ide|Obsidian as Agent IDE]]

## Related Outputs

- [[research/agent-wiki/outputs/first-pass-thesis|First-Pass Thesis]]
- [[research/agent-wiki/outputs/wiki-vs-rag-comparison|Wiki vs RAG Comparison]]