Note Content (Markdown) entities: - - 'Easier Hermes Infrastructure Plan: PERFECT FINAL VERSION' - Page topics: - EASIER_HERMES_INFRASTRUCTURE_PLAN - knowledge # Easier Hermes Infrastructure Plan: PERFECT FINAL VERSION **Status**: Approved & Active **Last Updated**: 2026-05-29 **Version**: 1.0.0 **Prepared By**: Hermes Agent (following Anthony Stratton's guidance) **Reference**: https://github.com/garrytan/gbrain + Easier Documentation --- ## Executive Summary This plan outlines a rock-solid infrastructure-first approach for the Easier Hermes instance, prioritizing knowledge base robustness before adding agents and workflows. It synthesizes: 1. **Karpathy's LLM Wiki pattern** (via Hermes' official `llm-wiki` skill) 2. **Garry Tan's gbrain advanced architecture** (production-tested brain system) 3. **Easier's documented best practices** from research and deployment guides 4. **Anthony Stratton's specific guidance** on infrastructure prioritization **Core Philosophy**: Build a self-maintaining knowledge compound where the agent handles maintenance overhead, leaving humans to curate sources, direct analysis, and ask strategic questions. --- ## Phase 0: Foundation Infrastructure (IMMEDIATE FOCUS) ### Goal: Establish auditable, self-maintaining knowledge infrastructure #### 0.1 Core Knowledge Base Architecture (Adopt gbrain structure) Following gbrain's proven directory structure, implement: ``` /easier-brain/ ├── AGENTS.md # Hermes rules and permissions (reviewed) ├── SCHEMA.md # Metadata and maintenance rules ├── index.md # Content catalog with one-line summaries ├── log.md # Chronological knowledge-maintenance log ├── RESOLVER.md # Master decision tree for filing (AGENT READS THIS FIRST) ├── schema.md # Page conventions, templates, workflows │ ├── people/ # One page per human being │ ├── README.md # Resolver: what goes here, what doesn't │ └── .raw/ # Raw API responses per person (JSON sidecars) │ ├── organisations/ # Client, prospect, partner, supplier pages │ ├── README.md │ └── .raw/ │ ├── projects/ # Easier Agency, Easier Now, tools, client projects │ └── README.md │ ├── decisions/ # Decision record plus why and evidence │ └── README.md │ ├── experiments/ # Hypothesis -> test -> measure -> outcome │ └── README.md │ ├── playbooks/ # Delivery and operational procedures │ └── README.md │ ├── concepts/ # Insights and frameworks │ └── README.md │ ├── briefs/ # Prepared outputs, not source evidence │ └── README.md │ ├── reviews/ # Daily/weekly/monthly synthesis and audits │ └── README.md │ ├── raw/ # Immutable evidence (append-only) │ ├── meetings/ │ ├── research/ │ ├── operations/ │ └── assets/ │ ├── receipts/ # Source attribution links │ └── README.md │ ├── digests/ # Processed summaries (Fathom, meeting extracts) │ └── README.md │ └── _archive/ # Historical reference only ``` #### 0.2 Critical Agent Configuration Rules Add these as **HARD RULES** in Hermes agent configuration (not suggestions): 1. **Before creating any brain page → read RESOLVER.md** (non-negotiable) 2. **Before answering any question about people, companies, deals, or strategy → search the brain first** 3. **The enrichment skill fires on EVERY signal** (meetings, email, social media, contacts) 4. **Corrections are highest-value data** → written to brain immediately, no batching 5. **Never modify raw/ evidence** – append-only, immutable source of truth #### 0.3 Implementation Stack (Phase 0 Recommended) Following Easier research documentation: - **Primary**: Hermes official `llm-wiki` skill (Karpathy pattern, official, simple) - **Evaluation**: QMD hybrid search (after benchmark confirms improvement on Easier questions) - **Delay**: gbrain installation until resource needs are clear on current VM - **Inspiration**: Use gbrain design principles immediately (graph links, gap analysis, dream-cycle) **VM Constraints Acknowledgement**: - Current host: e2-medium (4GB RAM, shared with n8n, Easier Done, Easier Marketing, nginx) - Do NOT run local embedding/model indexing until disk/memory headroom improves - Use gbrain's design as inspiration, treat installation as separate benchmarked decision #### 0.4 Zero Trust Initial State Per Easier research recommendations: - No messaging gateway - No API published to internet or private network - No n8n, Notion, CRM, ad-platform or email MCP integration - No autonomous curator, cron jobs, or automatic mass ingestion - No secrets in prepared configuration **Practical Target**: Progressive operating memory that first helps Anthony recall/synthesize/decide → then drafts/routes work for approval → eventually executes bounded, reversible operations with logs/budgets. --- ## Phase 1: Personal Research Vault ### Goal: Validate system with public research and self-written notes only #### Activities: - Run Hermes interactively with official `llm-wiki` skill enabled - Compare native retrieval with QMD hybrid search (after evaluation) - Perform weekly human-reviewed lint/audit: - Contradictions between pages - Stale claims superseded by newer sources - Orphan pages with no inbound links - Missing cross-references - Important concepts lacking dedicated pages - Process sources ONE AT A TIME with active supervision (preferred) - Build index.md and log.md through regular operations #### Exit Gate: - Retrieval quality test passes (correct evidence in top 5 for benchmark questions) - All generated claims properly cited with evidence - Knowledge base shows compounding value (less re-derivation over time) - Agent can say "not enough evidence" when appropriate --- ## Phase 2: Internal Business Memory ### Goal: Integrate approved non-client-sensitive Easier operational knowledge #### Activities: - Bring in approved strategy, product research, and operating notes - Begin approval queue for drafted decisions and tasks - Establish sensitivity labeling and access boundaries: - `internal` (default) - `client-confidential` - `personal-sensitive` - Test backup and restoration procedures (encrypted preferred) - Implement correction workflow for outdated information - Wire up enrichment to fire on every new person/company signal #### Exit Gate: - Backup, sensitivity rules, and correction workflow tested - System properly handles confidential vs. public information - Approval process for knowledge updates functioning - No sensitive data leakage inappropriately --- ## Phase 3: Confidential Material Integration ### Goal: Carefully introduce customer/prospect/meeting material under strict policies #### Activities: - Import selected material under retention, access, and deletion policies - Consider separate vault partitions for confidential client data - Implement strict provenance tracking for ALL external sources: - Source URL/date or import origin - Body checksum for change detection - Confidence/contested flags for multi-source claims - Maintain raw/ separation (immutable evidence never rewritten by agent) - Implement write-time retrieval: check existing memory before writing new compiled pages #### Exit Gate: - Privacy review passed - Tested access boundaries - Zero inappropriate sensitive data exposure - Provenance chain intact for all knowledge claims --- ## Phase 4: n8n and Easier Now Integration ### Goal: Expose narrowly scoped read or draft actions via controlled channels #### Activities: - Expose read/draft actions via n8n/Easier Now - **Human approval remains MANDATORY** for: - Messages to customers/prospects - Spend, legal, finance, credentials or access changes - Client deliverables - Data export - Workflow mutation - Build approval queues for drafted work (n8n or Easier Now mediated) - Maintain audit trails for all proposed actions - Implement operational disciplines: - Signal detection (brain-first lookup) - Sync-after-write (persistence guarantees) - Heartbeat monitoring - Dream cycle maintenance (when resource-appropriate) #### Exit Gate: - Per-workflow audit, undo plan, monitoring, and budget control - Clear separation between draft proposals and executed actions - All external actions flow through approval gates - System earns autonomy through reliable recall and auditability --- ## Phase 5: Controlled Autonomy (LATER) ### Goal: Allow narrow, reversible recurring actions #### Possible Automations (AFTER infrastructure validated): - Daily founder brief (draft-only until approved) - Weekly relationship risk review (draft until approved) - Weekly SOP improvement proposals - Research digest - Slack reminders to Anthony only (internal channel) #### Still Gated (PERMANENTLY): - Sending customer/prospect messages - Changing spend, ads, billing or contracts - Editing source-of-truth client records - Deleting or exporting private data - Autonomous actions without approval trails --- ## Key Operational Principles (From gbrain & Easier Research) ### 1. The Brain-Agent Loop (Core Compounding Mechanism) ``` Human & Approved Captures ↓ raw/ immutable evidence ────────────┐ │ │ │ controlled compilation │ retrieve exact evidence ↓ │ compiled Obsidian vault ◄───────────┘ │ │ relevant cited context only ↓ Hermes Agent (skills: llm-wiki, approved workflows) │ │ logs/checkpoints, no broad write/actions initially ↓ Approval queue (later via n8n / Easier Now) │ ↓ External action ONLY after bounded authorization ``` ### 2. Three Founding Principles (Karpathy/gbrain) **Principle 1: MECE Directories** Every piece of knowledge passes through RESOLVER.md decision tree and lands in exactly one directory. Prevents knowledge rot from duplicate pages with conflicting versions. **Principle 2: Compiled Truth + Timeline** - **Above the line**: Always current synthesis (rewritten when new info arrives) - **Below the line**: Append-only evidence log (never rewritten) → Enables instant current-state questions vs. historical investigation **Principle 3: Enrichment Fires on Every Signal** Every touchpoint (meeting, email, tweet, calendar, contact) automatically triggers enrichment. Knowledge grows as side effect of normal operations. ### 3. Entity Identity & Deduplication Protocol (Critical) Before creating ANY new page, agent MUST: 1. Search existing pages by name (exact and fuzzy) 2. Search aliases across all pages 3. Check .raw/ sidecars for matching email/social handles 4. **IF match found → UPDATE existing page** (add alias if new variant) 5. **IF no match → CREATE new page** ### 4. Write-Time Retrieval Discipline Before proposing new compiled page or strong claim: 1. Search for existing related pages and raw evidence 2. If nearly duplicate → LINK evidence to existing page 3. If updating existing belief → PRESERVE old evidence, MARK change 4. If contradictory → RETAIN BOTH claims, ROUTE for review 5. Write log entry describing proposal/update ### 5. Maintenance Discipline (Weekly Lint) Agent should regularly: - **Deduplication scan**: Similar names, same company, same email across pages - **Contradictions**: Conflicting facts between pages (different roles, etc.) - **Staleness**: State sections superseded by newer timeline entries - **Orphans**: Pages with no inbound links - **Open Threads**: Items seeming resolved but not moved to timeline - **Missing cross-references**: Entity A mentions B but doesn't link - **Missing pages**: Frequently mentioned entities lacking dedicated page - **MECE filing**: Pages that seem in wrong directory - **Source audit**: Unsourced claims in high-value sections (Beliefs, Motivations) - **Alias coverage**: Recent transcripts/emails with name variants not in aliases --- ## Technical Implementation Notes ### Hermes-Specific Configuration In `~/.hermes/config.yaml`: ```yaml # Enable core skills for knowledge base skills: llm-wiki: true # Add others incrementally after validation # Memory configuration - USE SPARINGLY memory: memory_enabled: true user_profile_enabled: true # DO NOT overload with business facts - use vault instead # Auxiliary model configuration - BE CONSERVATIVE ON CURRENT VM auxiliary: vision: provider: openrouter # or your preferred model: google/gemini-flash-1.5 # Delay compression/search until benchmarked ``` ### Required Toolsets for Knowledge Work ```bash hermes tools enable file search skills memory session_search web ``` ### Backup Strategy (Follow Easier Research) - Encrypted backups or encrypted git for versioned portability - Store large attachments outside vault (keep only receipts/digests/links in Hermes) - Monitor memory/disk/CPU before enabling QMD (2GB local model cited) - Use snapshot/disk policy before ingesting private documents --- ## Success Metrics & Exit Gates ### Phase 0 Completion: - [ ] RESOLVER.md created and referenced in agent config - [ ] Directory structure implemented per specification - [ ] Agent demonstrates MECE filing discipline - [ ] All new knowledge flows through enrichment pipeline ### Phase 1 Completion: - [ ] 30-50 source documents processed with supervision - [ ] Retrieval benchmark: 80%+ accuracy on Easier-style questions - [ ] Zero uncited claims in agent outputs - [ ] Weekly lint shows decreasing maintenance overhead over time ### Phase 2 Completion: - [ ] Approved internal Easier knowledge integrated - [ ] Sensitivity labeling system functional - [ ] Backup/restore tested and documented - [ ] Approval queue for drafted work operational ### Phase 3 Completion: - [ ] Confidential material imported under strict policy - [ ] Provenance tracking verified for all knowledge - [ ] Zero inappropriate data exposure in testing - [ ] Access boundaries validated ### Overall Success: System earns trust through: - Memory quality (accurate, cited, current) - Clear roles (human curates, agent maintains) - Reviewable work (all changes in log.md with provenance) - Reversible automation (no irreversible actions without approval) --- ## Immediate Next Steps (TODAY) 1. **Create RESOLVER.md** based on gbrain's master decision tree 2. **Implement directory structure** in `/srv/easier-hermes/vault/` 3. **Configure Hermes agent** to reference RESORLER.md as hard rule 4. **Run initial llm-wiki skill** to verify basic functionality 5. **Create benchmark question set** for memory layer evaluation 6. **Begin supervised source processing** (one document at a time) > **Remember**: The human's job is to curate sources, direct analysis, ask good questions, and think about what it all means. > The agent's job is everything else: summarizing, cross-referencing, filing, and bookkeeping that makes a knowledge base actually useful over time. This infrastructure-first approach ensures we build a rock-solid foundation that will compound value over time, rather than accumulating technical debt that must be repaid later. --- *This document serves as the authoritative reference for Easier Hermes infrastructure development. Refer to it before making architectural decisions or adding complexity to the system.*

Save Changes