---
type: output
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/amp-how-to-build-an-agent]]"
  - "[[research/agent-wiki/sources/dissenting-views-on-persistent-agent-wikis]]"
derived_from:
  - "[[research/agent-wiki/concepts/compiled-knowledge-layer]]"
  - "[[research/agent-wiki/concepts/simplicity-before-autonomy]]"
  - "[[research/agent-wiki/concepts/obsidian-as-agent-ide]]"
  - "[[research/agent-wiki/concepts/governed-interpretation-layer]]"
confidence: medium
output_kind: memo
---

# First-Pass Thesis

## Question

What does this workspace currently believe about LLM-maintained knowledge systems and agent design?

## Answer

The workspace currently believes that the most promising form of agent memory for this vault is not a giant retrieval layer over all notes, but a bounded, maintained markdown workspace that compounds useful synthesis over time.

The Karpathy pattern is directionally right because it treats the wiki as the durable artifact rather than treating every question as a fresh retrieval problem. But the Anthropic and Amp sources add an important correction: the system should stay simple, inspectable, and narrow in scope. A good agent-wiki system is not "maximum autonomy." It is a small number of explicit operations over a bounded corpus, with the human staying in charge of scoping, thesis formation, and promotion.

The new dissenting source sharpens that thesis further. The workspace should not treat the compiled layer as silent authority. Persistent synthesis is useful, but it also creates delayed attack surfaces, provenance thinning, and drift risk if the wiki is allowed to replace verification entirely. The correct shape is therefore hybrid: compiled knowledge for continuity and reusable interpretation, retrieval over canonical notes for verification, recency, and high-stakes factual grounding.

This implies a practical architecture for the vault:

- canonical source notes remain read-only
- `research/<topic>/` is the active compiler-owned workspace
- concepts are local first
- shared concepts are promoted only when reuse is earned
- outputs are saved selectively when they add durable value

In other words, the workspace is currently converging on a philosophy of compiled knowledge with restrained autonomy.

## Basis

- [[research/agent-wiki/sources/karpathy-llm-wiki]] provides the core compiled-wiki idea.
- [[research/agent-wiki/sources/anthropic-building-effective-ai-agents]] provides the "start simple" constraint.
- [[research/agent-wiki/sources/amp-how-to-build-an-agent]] grounds the system in implementation realism.
- [[research/agent-wiki/sources/dissenting-views-on-persistent-agent-wikis]] provides the trust boundary: interpretation should compound, but truth still needs a backstop.

## Follow-On Questions

- What is the lightest possible daily or weekly ritual that keeps this workspace alive without weakening provenance?
- Which future implementation case studies would test whether this hybrid model actually scales?
- Which concept should be the first serious candidate for promotion into `knowledge/concepts/` once it appears outside this workspace?