Overview
mycelium-hq/ai-brain-starter is a verification harness for Claude Code memory (26⭐). It is listed in the vault’s Resources page as a tool/repository — focused on verifying that agent memory is correct.
Key Facts
- Stars: ~31 (as of 2026-08-02)
- Purpose: verification harness for Claude Code memory
- Role: tests/validates agent memory correctness
- Framework: 4 reliability mechanisms (Markdown ground truth, deterministic hooks, bi-temporal rule lineage, session routines)
- License: MIT
Features
Core Architecture
- Vault as ground truth — markdown on disk as the authoritative source
- Bi-temporal rule lineage — rules carry written/verified timestamps; drift detection scans high-edit files
- Session lifecycle — start loads context; end cascades decisions/insights/to-dos to right files
- 40+ skills — comprehensive skill library for agent operations
- 15 accountability rules — embedded in CLAUDE.md
Ingest & Processing
/journal— daily journal interview with advisory panel, emotional floor tracking/meeting-todos— pull action items from meeting notes/patterns— extract recurring patterns into permanent captures/deconstruct— first-principles analysis
Search & Retrieval
/mem-search— cross-session memory search/graphify— knowledge graph extraction with community detection
Quality & Governance
- Deterministic hooks — pre-write hooks block bad edits (missing fields, formatting issues)
- Drift detection — flags files edited 5+ times in 30 days
- Rule conflict detection — catches cross-document contradictions at write time
- Cited agent actions — all memory operations traceable
MCP / Tools
- 8 plugin marketplaces — Trail of Bits, Stripe, Cloudflare, Sentry, Obsidian, etc.
- 13 companion MCP servers — Slack, iMessage, WhatsApp, Google Workspace, etc.
- MCP servers — Granola (meeting transcription), ChatPRD (PRD drafting)
/optimize-brain— deep vault optimization
Integration
- Claude Code native — designed for Claude Code memory verification
- Obsidian integration — Templater templates, vault structure
- Dataview + Bases — live database queries over markdown
- Personal-to-team path — bidirectional team folder between personal/team vaults
- Corporate install profile — hardened, telemetry-off, pinned versions
- One-paste install —
git clone+bootstrap.sh
Quality & Governance
/coaching— multi-pass coaching session with accountability record/weekly— weekly pattern recognition across journal entries/monthly— deeper monthly trends/humanizer— de-AI-ify writing, score against voice corpus- Weekly/monthly reports — pattern tracking, avoidance flagging
Data & Storage
- Markdown files — plain-text versioned memory
- Audit logs — track all memory operations
- Tenant/role boundaries — access control for team use
Relationships
- “depends-on” claude-code — verifies its memory
- “belongs-to” agent-memory-systems — memory verification
- “sibling-of” lint — complementary to the health-check pass
- “sibling-of” memex-zero-rag — both enforce correctness (quarantine versus verification)
Product and Governance Layer
Mycelium’s product page positions this open-core memory layer beneath role-scoped AI teammates. The described architecture connects admin-approved systems, applies tenant and role boundaries to memory, and returns cited agent actions. It highlights customer-owned deployment, plain-text versioned memory, audit logs, independent read/write/synthesis permissions, and 23 live connectors.
Implications: The repository is presented as both a personal verification harness and the substrate for governed team memory, extending the LLM Wiki pattern from individual context to controlled organizational workflows.
Reliability Harness
The repository describes four reliability mechanisms: Markdown on disk as ground truth, deterministic hooks that block malformed or unsafe edits, bi-temporal rule lineage with verification and drift detection, and session-start/session-end routines that move decisions, insights, and to-dos out of chat and into the vault. It also describes a personal-to-team path with separate team-vault boundaries and a shared-brain model.
The repository reports a measured-sounding claim that professional content can be corrupted during repeated model edits, but the captured page presents that claim as project rationale rather than independent validation. The earlier vault estimate of approximately 26 stars and the captured page’s 31 stars are time-separated observations, not a substantive contradiction.
Implications: AI Brain Starter extends memory retention into memory verification: the system attempts to prevent corruption before it becomes durable and to reconcile context across sessions.
Real-World Adoption Evidence
Adelaida Diaz-Roa deployed the ai-brain-starter skill (via myceliumai.co/install) as part of her integrated second-brain system, measuring setup time reduction from approximately one week to “approximately one afternoon”—a concrete, time-bound adoption signal. She reports using the tool to automate setup of vault initialization through guided conversation, reducing implementation friction at the entry point. See adelaida-diaz-roa-second-brain-knowledge for the filed query on what kind of knowledge her second brain accumulates.
Implications: The skill is not merely a technical tool but a friction-reduction gateway. Early adoption by knowledge-management practitioners suggests the tool solves a real implementation pain point—bootstrapping a new personal vault in minutes rather than days.
Implications
A verification harness is the missing piece for trust: it checks the agent’s memory, much like this vault’s wikilint checks link health. Relevant if the user wants automated correctness guarantees. The real-world adoption evidence suggests this tool bridges the gap between the conceptual appeal of LLM Wikis and the practical effort required to implement them.
External Links
- Repository: https://github.com/mycelium-hq/ai-brain-starter
- Installation: https://myceliumai.co/install
Sources
^[wiki/concepts/resources.md] ^[raw/external/mycelium-ai-co-25c90bd6.md]