Building a Second Brain for Claude Code: Hermes-Inspired Personalization

The Goal

#

Turn Claude Code from a stateless coding assistant into a persistent personal assistant — a second brain that remembers context across sessions, learns from daily work, and can semantically search past knowledge.

Inspired by Hermes’s personalization architecture, adapted for Claude Code’s native capabilities.

Hermes’s Approach (The Model)

#

Hermes creates personalization through:

Gap Analysis: Claude Code vs Hermes

#

Claude Code already has ~70% of the infrastructure. The gap is automation glue, not features.

Architecture

#

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

LAYER 1: CURATED MEMORY (auto-loaded every session)
└── ~/.claude/MEMORY.md (first 200 lines)
    ├── Identity & Preferences
    ├── Active Projects
    ├── Patterns & Decisions
    └── Recent Context (rotated weekly)

LAYER 2: DAILY CHRONICLE (hook-automated)
└── ~/.claude/memory/YYYY-MM-DD.md
    ├── Session summaries
    ├── Decisions made
    ├── Problems encountered
    └── Follow-ups needed
    Hooks: PreCompact → flush, SessionStop → summarize

LAYER 3: SEARCHABLE KNOWLEDGE (MCP server)
└── ChromaDB + bge-small-en-v1.5
    ├── Indexes all memory/*.md files
    ├── Semantic vector search (cosine similarity)
    └── Nightly re-index via cron

LAYER 4: LIVING SKILLS (on-demand)
└── /recall, /journal, /reflect skills
    ├── Search memory corpus
    ├── Create structured daily entries
    └── Weekly synthesis + promotion to MEMORY.md

Implementation Plan

#

Phase 1: Quick Wins (1-2 hours)

#

• Restructure MEMORY.md — Curated 200-line template with sections: Identity, Active Projects, Preferences, Architectural Decisions, Patterns Learned, Recent Context
• Create PREFERENCES.md — Centralized personal preferences (communication style, tech choices, work habits)
• Create daily log system — ~/.claude/memory/ directory with .template.md
• Create PreCompact hook — memory-flush.sh writes context to daily log before compaction
• Create SessionStart hook — session-start-memory.sh loads today’s log + yesterday’s follow-ups
• Create SessionEnd hook — session-summary.sh appends session boundary to daily log
• Configure hooks — Merge into settings.local.json
• Activate OMC notepad — Add reminder to use addLearning(), addDecision() during planning

Phase 2: Enhanced Memory (3-4 hours)

#

• Weekly reflection script — weekly-reflect.sh synthesizes 7 days, flags promotion candidates (cron Sunday 9pm)
• Memory cleanup script — memory-cleanup.sh archives logs >30 days as .gz (cron monthly)
• Create /recall skill — Search across all memory files (daily logs, weekly syntheses, project memories)
• Create /journal skill — Structured daily entry creation with type classification
• Create /reflect skill — Weekly review and promotion to MEMORY.md (keeps under 200 lines)
• Set up cron jobs — Weekly reflection + monthly cleanup + nightly reindex

Phase 3: Vector Search Brain (4-6 hours)

#

• Build ChromaDB MCP server — Python MCP service with bge-small-en-v1.5 embeddings for semantic search
• Register MCP server — Added to ~/GIT/.mcp.json (workspace-level)
• Set up nightly reindex — Cron job at 11pm Pacific to keep vectors fresh
• Optional: Obsidian bridge — Add vault path to indexer for personal notes

Approach Comparison

#

Why Full Hybrid?

#

1. Phase 1 is zero-risk — hooks + markdown that make every session smarter, starting immediately
2. Phase 2 makes it self-sustaining — weekly synthesis prevents memory rot
3. Phase 3 is the real unlock — within 3 months there will be 90+ daily logs; without semantic search, old memories become unfindable

Plugins (Claude-Mem, Supermemory) are designed for users without sophisticated setups. They would conflict with the existing oh-my-claudecode orchestration layer and some require cloud APIs.

Key Design Decisions

#

• ChromaDB over sqlite-vec/Qdrant/Pinecone — Industry default for local AI apps (25K GitHub stars, used by Claude-Mem plugin). Built-in HNSW index, PersistentClient for on-disk storage, cosine distance. sqlite-vec is lighter but less community support and tooling.
• bge-small-en-v1.5 over all-MiniLM-L6-v2 — Same 384 dimensions but 62% MTEB score vs 28%. Released 2023 vs 2019. Drop-in replacement with dramatically better retrieval quality.
• 200-line MEMORY.md budget — Claude Code auto-loads first 200 lines; every line must earn its place
• Daily logs are append-only — Never edit past entries, only promote insights upward
• Weekly reflection is manual — Script generates candidates, human approves promotions

Lessons Learned

#

• Hermes’s power isn’t any single feature — it’s the lifecycle: capture daily, search semantically, synthesize weekly, promote durably
• Claude Code already has 70% of the infrastructure; the gap is automation glue
• The 200-line MEMORY.md limit is actually a feature — forces curation over accumulation
• Hooks (PreCompact, SessionStart, SessionEnd) are the critical integration points — note: SessionStop is not a valid hook event, use SessionEnd
• Local-first vector search (ChromaDB + bge-small-en-v1.5) beats cloud solutions for privacy and reliability
• Embedding model choice matters: bge-small-en-v1.5 (2023, 62% MTEB) dramatically outperforms all-MiniLM-L6-v2 (2019, 28% MTEB) at the same 384 dimensions