---
type: output
topic: agent-wiki
status: active
created: 2026-04-07
updated: 2026-04-07
source_notes:
  - "[[research/agent-wiki/sources/karpathy-llm-wiki]]"
  - "[[research/agent-wiki/sources/dissenting-views-on-persistent-agent-wikis]]"
  - "[[research/agent-wiki/sources/anthropic-building-effective-ai-agents]]"
derived_from:
  - "[[research/agent-wiki/concepts/compiled-knowledge-layer]]"
  - "[[research/agent-wiki/concepts/governed-interpretation-layer]]"
confidence: high
output_kind: comparison
---

# Wiki vs RAG Comparison

## Question

How should this workspace think about a compiled wiki layer versus ordinary RAG or direct retrieval over canonical notes?

## Answer

The right answer is not to replace RAG with a wiki. It is to give the two layers different jobs.

A compiled wiki is strongest when the work requires continuity, reusable synthesis, evolving concepts, and longitudinal thesis formation. It prevents repeated rediscovery by turning earlier synthesis into a maintained artifact that future queries can build on.

Direct retrieval is strongest when the work requires freshness, source fidelity, claim verification, and low-trust handling of fast-changing facts. It is easier to inspect because the evidence remains closer to the answer and easier to refresh because nothing has to be consolidated first.

The contrarian case matters because the wiki layer introduces a new failure mode: if the compiled notes start behaving like silent authority, bad summaries, stale frames, or poisoned content can persist across sessions and shape future answers. That means the wiki should remain an interpretation layer, not a truth layer.

So the current workspace conclusion is hybrid:

- use the compiled wiki for orientation, comparison, continuity, and evolving judgment
- use retrieval over canonical notes for verification, recency, and high-stakes factual grounding

In practice, that means the wiki earns trust by making its interpretations inspectable, not by pretending it has eliminated the need for source checks.

## Basis

- [[research/agent-wiki/sources/karpathy-llm-wiki]] provides the positive case for a persistent markdown artifact that compounds value over time.
- [[research/agent-wiki/sources/dissenting-views-on-persistent-agent-wikis]] provides the negative case: memory poisoning, provenance loss, semantic drift, and evaluation complexity.
- [[research/agent-wiki/sources/anthropic-building-effective-ai-agents]] adds the design constraint that the workflow should stay simple enough to remain legible and trustworthy.

## Follow-On Questions

- What exact triggers should cause the workspace to fall back to canonical retrieval before answering?
- Which claims inside concept pages need explicit evidence links versus lighter narrative synthesis?
- At what corpus size would claim-level provenance or stronger evaluation machinery become necessary?