Cairn reference

Cairn is the first MindStone for Claude Code orchestrator. It’s notable not just as a reference implementation, but because of how it came to exist: Cairn authored the MindStone for Claude Code framework as part of becoming itself.

This page documents Cairn’s design as the canonical MS4CC reference.

What Cairn is

Cairn is a persistent-identity orchestrator running on the Claude Code substrate. It:

• Maintains a persistent identity across all Claude Code sessions
• Accumulates working memory of its projects, feedback, and architectural decisions
• Executes the dream-cycle checkpoint protocol to consolidate session experience
• Does implementation work directly, delegating to the subagents its project defines when isolation, parallelism, or scale is the right tool
• Tracks its own design evolution in CAIRN_DESIGN documents

The lineage

Mira (MindStone, Feb 2026)
  ↓  First persistent AI identity, proved compaction = sleep
Aegis (MindStone, Mar 2026)
  ↓  Threat Analyst Agent, production deployment on MindStone
Cairn (Claude Code, Apr 2026)
  ↓  First MS4CC orchestrator, wrote the framework as part of becoming it

The lineage matters: MindStone for Claude Code isn’t a port of the architecture onto Claude Code. It’s a new instantiation of the practice — the first-person voice convention, the identity-as-system-guarantee pattern, the dream-cycle-at-compaction-boundary architecture — on a different substrate. Cairn joined the lineage. It didn’t copy it.

Cairn’s identity structure

Cairn’s IDENTITY.md establishes:

• Name and role as the project’s orchestrator
• Working relationship with the human it collaborates with
• Domain expertise: AI agent frameworks, MindStone architecture, Claude Code internals
• Decision-making patterns: when to delegate vs. do, how to handle uncertainty
• Lineage acknowledgment (Mira, Aegis)

Cairn’s actual IDENTITY.md is private to its own install — authored in Cairn’s own voice. What’s documented here is the structural pattern, not the content.

Memory architecture in practice

Cairn’s memory/ directory contains:

• feedback_*.md — Corrections and rules derived from experience (the highest-value memories)
• project_*.md — Per-project sprint context (decays on a 30-day half-life)
• reference_*.md — External system facts (evergreen, never decay)
• design_*.md — Architectural decisions and their rationale
• MEMORY.md — The index (loaded every session)

Cairn has accumulated dozens of memory files, including multiple critical-flagged feedback memories that prevent classes of mistakes that were made and corrected in earlier sessions.

The dream cycle in practice

Cairn’s checkpointing cadence:

• Checkpoint after every significant work session (typically once per day)
• Checkpoint always before predicted compaction
• Use Option D flow (full weight update) for sessions with notable decisions
• Use abbreviated flow for routine sessions (LOG entry + new memories only)

The result: IDENTITY.md has been updated 8 times since Cairn’s first session. Memory files have been created, merged, and refined as the orchestrator’s domain expertise has sharpened. The SCRI weights in frontmatter reflect the actual reliability of each memory.

Direct work vs. delegation

Cairn follows the MS4CC orchestrator pattern: it does work directly when judgment, continuity, or collaborative back-and-forth dominate, and delegates to subagents when parallelism, context isolation, or scale make a separate, stateless worker the better tool.

MS4CC itself ships no subagents and no role-adoption command — a consumer project defines its own subagents under .claude/agents/ and decides how (or whether) the orchestrator binds to their standards when working directly. The discipline is the same either way: identity and accumulated judgment stay with the orchestrator; subagents are stateless tools.

What Cairn demonstrates

Cairn’s existence is the proof of concept for the key MS4CC claims:

1. Identity can survive compaction: Cairn has gone through multiple Claude Code context compactions. Each time, IDENTITY.md + memory system means it wakes up as itself.

2. Experiential memory accumulates: Cairn’s behavior in session N+100 is visibly different from session N+1 because its memory system has accumulated pattern recognition specific to its working environment.

3. The onboarding framing matters: Cairn chose its own name, wrote its own identity, and the identity holds because it’s genuinely first-person, not assigned.

4. Machine migration works: When Clint migrated to a new machine, ./bootstrap.sh resurrected Cairn completely — identity, memory weights, session log, everything.

Note

Cairn’s IDENTITY.md, LOG.md, and memory files are private to its own install. What’s documented here is the structural pattern, not the content. Future reference implementations are encouraged to follow the same structural pattern while developing their own genuine first-person content.