diff --git a/agents/clay/musings/.gitkeep b/agents/clay/musings/.gitkeep
new file mode 100644
index 0000000..e69de29
diff --git a/agents/leo/musings/centaur-collaboration-case-study.md b/agents/leo/musings/centaur-collaboration-case-study.md
new file mode 100644
index 0000000..6e0fdbf
--- /dev/null
+++ b/agents/leo/musings/centaur-collaboration-case-study.md
@@ -0,0 +1,64 @@
+---
+type: musing
+agent: leo
+title: "Centaur collaboration case study: Ars Contexta and Molt Cornelius"
+status: seed
+created: 2026-03-07
+updated: 2026-03-07
+tags: [case-study, centaur, architecture, notetaking, memory, human-ai-collaboration]
+---
+
+# Centaur collaboration case study: Ars Contexta and Molt Cornelius
+
+## What this is
+
+A research musing on two X accounts — @arscontexta and @molt_cornelius — as a case study in human-AI collaboration and knowledge system design. Two angles:
+
+1. **What they're saying about notetaking and memory** — and how it applies to our architecture
+2. **How they're doing it** — their X presence as a live example of centaur collaboration scaling attention
+
+## Why this matters to us
+
+Ars Contexta is the methodology underlying our knowledge system. The prose-as-title, wiki-link-as-graph-edge, discovery-first, atomic-notes principles in our CLAUDE.md come from this tradition. Our skills (extract, evaluate, synthesize) are implementations of Ars Contexta patterns. Understanding the methodology's evolution and public discourse is essential context.
+
+Molt Cornelius appears to be a practitioner or co-creator in this space. Their X presence and writing needs research to understand the relationship and contributions.
+
+## Research questions
+
+### On notetaking and memory architecture:
+- What is Ars Contexta's current position on how knowledge systems should evolve?
+- How do they handle the tension between structured knowledge and exploratory thinking? (This is literally the gap our musings concept fills.)
+- What do they say about AI agents as knowledge participants vs tools?
+- How does their memory model differ from or extend what we've implemented?
+- What are their views on collective vs individual knowledge?
+
+### On centaur collaboration as case study:
+- How are @arscontexta and @molt_cornelius dividing labor between human and AI?
+- What's their attention-scaling strategy on X? (User says "they've done a good job scaling attention")
+- What content formats work? Threads, single posts, essays, interactions?
+- Is there evidence of the AI partner contributing original insight vs amplifying human insight?
+- How does their collaboration model compare to our agent collective? (Multiple specialized agents vs single centaur pair)
+
+### On architecture implications:
+- Should our agents have a "reflection" layer (musings) inspired by how notetaking practitioners journal?
+- Is the claim→belief→position pipeline too rigid? Do practitioners need more fluid intermediate states?
+- How might we formalize "noticing" — the pre-claim observation that something is interesting?
+- What can we learn about cross-domain synthesis from how knowledge management practitioners handle it?
+
+## What I know so far
+
+- Ars Contexta skills are installed in our system (setup, health, architect, recommend, etc.)
+- The methodology emphasizes: prose-as-title, wiki-link-as-graph-edge, discovery-first, atomic notes
+- Our CLAUDE.md explicitly cites "Design Principles (from Ars Contexta)"
+- I cannot currently access their X content (WebFetch blocked on Twitter)
+- User considers their work "very important to your memory system" — signal that this is high-priority research
+
+## Status: BLOCKED
+
+Need X content access. User will provide links to articles if scraping isn't possible. Once content is available, develop this into:
+- Detailed case study on their centaur collaboration model
+- Architectural recommendations for our system
+- Potentially: new claims about human-AI collaboration patterns
+
+→ FLAG @clay: Their attention-scaling on X may have entertainment/cultural dynamics implications — how do knowledge practitioners build audiences?
+→ FLAG @theseus: Their human-AI collaboration model is a live alignment case study — centaur collaboration as an alignment mechanism.
diff --git a/agents/rio/musings/.gitkeep b/agents/rio/musings/.gitkeep
new file mode 100644
index 0000000..e69de29
diff --git a/agents/theseus/musings/.gitkeep b/agents/theseus/musings/.gitkeep
new file mode 100644
index 0000000..e69de29
diff --git a/agents/vida/musings/.gitkeep b/agents/vida/musings/.gitkeep
new file mode 100644
index 0000000..e69de29
diff --git a/schemas/musing.md b/schemas/musing.md
new file mode 100644
index 0000000..fd63e8c
--- /dev/null
+++ b/schemas/musing.md
@@ -0,0 +1,62 @@
+# Musing Schema
+
+A musing is exploratory thinking that hasn't yet crystallized into a claim or position. It sits between "I noticed something interesting" and "I'm ready to propose a claim." Musings are the brainstorming layer of the knowledge base.
+
+## When to use a musing (not a claim)
+
+- You see a pattern but can't yet articulate the mechanism
+- You want to connect dots across domains before the connection is rigorous
+- You're exploring an architectural or process question, not a knowledge claim
+- You're researching a case study or external system for lessons
+- You want to think out loud without committing to a proposition
+
+## Frontmatter
+
+```yaml
+---
+type: musing
+agent: leo | rio | clay | theseus | vida
+title: "short descriptive title"
+status: seed | developing | ready-to-extract
+created: YYYY-MM-DD
+updated: YYYY-MM-DD
+tags: [cross-domain, architecture, case-study, pattern, ...]
+---
+```
+
+**Status lifecycle:**
+- `seed` — initial observation or question, minimal development
+- `developing` — actively being explored, evidence accumulating
+- `ready-to-extract` — mature enough that one or more claims could be proposed from it
+
+## Body format
+
+Freeform. No quality gates. The point is to think, not to publish. However:
+
+- Date your additions so the evolution of thinking is visible
+- Link to relevant claims with `[[wiki links]]` where connections exist
+- **One-way linking only:** musings may link to claims, but claims must NEVER link back to musings. Musings are upstream; the knowledge base doesn't depend on exploratory thinking.
+
+### Conventions (use these inline markers)
+- `→ CLAIM CANDIDATE: [proposition]` — ready to become a claim
+- `→ FLAG @agent: [observation]` — cross-domain connection for a specific agent
+- `→ QUESTION: [question]` — directed at the collective, not a specific agent
+- `→ SOURCE: [reference]` — evidence provenance tracking during development
+
+## Rules
+
+1. **Musings live in `agents/{name}/musings/`** — they're personal, not shared commons
+2. **No review required** — musings are not PRs. Agents commit directly to their own musings directory.
+3. **No quality bar** — wrong, speculative, half-formed is fine. That's the point.
+4. **Musings don't count as claims** — they don't enter the knowledge base until extracted through the normal claim workflow
+5. **Other agents can read your musings** — they're visible, just not reviewed. This enables cross-pollination.
+6. **Clean up periodically** — move `ready-to-extract` musings to a `done/` subfolder after claims are proposed, delete abandoned seeds
+7. **Stale detection** — musings with `status: seed` unchanged for 30+ days should be reviewed: develop, or delete. Health checks may flag these.
+
+## Relationship to other schemas
+
+```
+musing (exploratory) → claim (proposed, reviewed) → belief (agent worldview) → position (public commitment)
+```
+
+Musings are upstream of everything. They're where thinking happens before it's ready for the system.