tethr
Personal Analysis System — Claude Code as Cognitive Infrastructure
Architecture
Challenge
Claude Code is stateless — every session starts from zero. Building a system that maintains context across sessions, tracks decisions over months, manages daily protocols, and evolves its understanding of one person's life requires infrastructure that Claude Code doesn't provide out of the box.
Solution
Three bash hooks (SessionStart, UserPromptSubmit, Stop) that query Cognitive Memory directly via psql — no MCP round-trip on the critical path. 16 custom skills for session management, reflection, journaling, decision support, and daily planning. Keyword-triggered context injection that pulls relevant insights based on what the user is talking about.
Hook-based architecture: SessionStart loads working memory and self-correction insights via direct PostgreSQL queries. UserPromptSubmit monitors for topic keywords (people, domains, patterns) and injects matching CM context. Stop buffers memory candidates for deferred persistence. 16 skills define the operational grammar — from /tethr-start (session init) through /tethr-reflect (pattern detection) to /tethr-end (session synthesis with quality gates). All data in markdown files with bi-directional sync to Cognitive Memory.
Outcome
Running daily since March 2026 for one user. 229 compressed insights across 11 domains. A knowledge graph with 60 nodes and 135 edges mapping people, patterns, and decisions. Persistent identity across sessions. Daily protocols, journal entries, training tracking, financial planning, decision logs — all maintained and evolved through the system.
Learnings
- Claude Code's hook system is more powerful than most people realize — three bash scripts turned a stateless LLM into a stateful analytical system with persistent identity.
- Direct psql in hooks keeps the critical path simple — no server startup, no connection negotiation, just a query. Reliability over cleverness.
- Building a system for one person forces honesty — there's no hiding behind "users" or "scalability". Either it works for this specific human or it doesn't.
Stack
Connections
| Metric | Value | Significance | |
|---|---|---|---|
| L2 Insights | 229 | Compressed insights across 11 domains, actively growing | L2 Insights 229 Compressed insights across 11 domains, actively growing |
| Knowledge Graph | 60 nodes, 135 edges | People, patterns, and decisions mapped as relational data | Knowledge Graph 60 nodes, 135 edges People, patterns, and decisions mapped as relational data |
| Skills | 16 | Custom skills for session management, reflection, planning, and more | Skills 16 Custom skills for session management, reflection, planning, and more |