The Amnesia inherent in a New Chat

Started: 2025-01-22

Summary of steps taken during "go ws" initialization, plus research on AI memory systems.

Steps

1. Attempted ws CLAUDE.MD — failed, wrong path (/Users/jonathanmast/...)
2. Listed allowed directories — discovered correct base is /Users/sand/GitHub
3. Explored mono structure — found mono/projects/ws/CLAUDE.MD
4. Read ws/CLAUDE.MD — instructions: read ../../CLAUDE.MD and notes/guides/*.md
5. Read mono/CLAUDE.MD — shared context: defaults, gates, workarounds
6. Read collaborate guides (batch 1) — index, access, chat, evolve, filesystem, gating
7. Read collaborate guides (batch 2) — journals, markdown, repo, shorthand, voice, workflow
8. Read mono/notes/work files — index, revisit.md, journal.md
9. Explored ws/notes — found collaborate dir (no guides dir)
10. Read ws-specific guides — gotchas.md, composition.md (Svelte patterns)
11. Read ws/notes/work files — index, deliverables.md (bug tracking)
12. Searched for recents docs — found done/recents.md
13. Read recents.md — learned Phase 5 complete, redesign finished Jan 21
14. Summarized state — noted memory about "Phase 3 next" is stale

Observations

• Path discovery took 4 tool calls (could be 1 if I knew the base path)
• ws/CLAUDE.MD says "Read all .md files in notes/guides/" but that dir doesn't exist — actual guides are in notes/collaborate/
• Memory said "Phase 3 next" for recents but the work is complete

userMemories Staleness Discovery

Problem: userMemories contained stale entry "ws recents redesign: Phase 1-2 done, Phase 3 next" — but that work completed Jan 21.

Root cause: userMemories updates "periodically in the background" — exact period unknown. System prompt says deletion cleanup is "nightly" but creation/update cadence is undocumented.

Freshness comparison:

Source	Freshness
CLAUDE.MD, guides, work/*.md	Fresh — read live
conversation_search / recent_chats	Fresh — searches actual history
userMemories (auto-generated)	Stale — unknown lag
memory_user_edits (manual)	Immediate, reliable

Resolution:

• Removed stale memory entry
• Updated CLAUDE.MD to say: "userMemories: Ignore entirely. Stale data risk outweighs serendipity. Use memory command if curious."
• Added stable context (Who section, execution mode) directly to CLAUDE.MD
• Added memory shorthand command to surface auto-generated insights on demand

Implication: Don't trust userMemories for anything time-sensitive. The md files are the real memory system.

AI Memory Systems Research

Searched for consumer products that accomplish something like Jonathan's md-based memory system.

Products Found

PAI (Personal AI Infrastructure) — Daniel Miessler, 5.4k GitHub stars

• TELOS: 10 identity files (MISSION.md, GOALS.md, BELIEFS.md, etc.)
• User/System separation (customizations survive upgrades)
• Skill system with routing
• Memory system with learning signals (ratings, sentiment)
• Hook system (lifecycle events)
• "Continuous learning" — captures feedback
• Philosophy: "AI as persistent assistant, friend, coach, mentor"
• Gap: No gates, no discipline enforcement, more "life coach" than project workflow

Basic Memory — basicmachines-co

• Local markdown files as knowledge graph
• Bidirectional read/write (human and AI edit same files)
• SQLite index for semantic search
• Obsidian-compatible
• True local-first, no cloud
• Gap: Storage system only, no workflow enforcement

Nick Cao's Memory Bank — 5-mode workflow MCP server

• .memory_bank directory with tasks.md, implementation-plan.md
• 5-mode lifecycle: VAN (init) → PLAN → CREATIVE → IMPLEMENT → REFLECT+ARCHIVE
• Complexity levels (1-4)
• File-based persistence
• Gap: Rigid phases, no organic evolution, no gates

Other tools: Cursor (.cursorrules), Claude Projects, Windsurf, Mem0, various MCP servers

Feature Comparison

Feature	PAI	Basic Memory	Nick Cao	Jonathan's System
Persistent context	✅	✅	✅	✅
Markdown-based	✅	✅	✅	✅
Work tracking	❌	❌	✅ (phases)	✅ (resume points)
Gates	❌	❌	❌	✅
Discipline enforcement	❌	❌	❌	✅
"Leaning into learning"	partial	❌	❌	✅
Shorthand commands	❌	❌	❌	✅
Voice/tone guide	❌	❌	❌	✅
Organic evolution	❌	❌	❌	✅

Unique Contributions of Jonathan's System

1. Gates — forced checkpoints before action (refactoring → read workflow.md first)
2. Discipline enforcement — recovery paths ("lazy" → reread refactoring discipline)
3. "Leaning into learning" — iterative guide refinement from actual collaboration
4. Shorthand as UI — compact commands for collaboration flow
5. Voice guide — personality/tone consistency in documentation

Key insight: Most tools provide storage and retrieval. Jonathan's system provides accountability and learning loops. It's methodology, not just tooling.

The Irony

Jonathan built an entire methodology — CLAUDE.MD, gates, work files, guides — to work around AI unreliability. That labor compensates for the tool's opacity. The collaboration guides he's written are clearer documentation of how to work with Claude than anything Anthropic provides.