andrem-sec Continuous Learning v2 Skill
Cross-session learning system that captures behavioral patterns as instincts with confidence scoring and automatic promotion to global scope.
What Claude Gets Wrong Without This Skill
Without continuous learning, Claude:
- Rewrites skills manually instead of learning from patterns
- Forgets project-specific patterns between sessions
- Cannot differentiate high-confidence vs low-confidence patterns
- Applies one-size-fits-all rules instead of context-specific instincts
- Has no path for local instincts to become global best practices
Continuous learning tracks what works, increases confidence with successful applications, and promotes proven patterns globally.
The Instinct System
Instinct: Atomic behavioral unit with trigger (when), action (what), and confidence (0.0-1.0).
Example:
{
"id": "inst_001",
"trigger": "writing test file",
"action": "colocate test next to source (src/user.ts -> src/user.test.ts)",
"confidence": 0.85,
"domain": "testing",
"evidence": ["2026-03-28: Created tests/user.test.ts next to src/user.ts"],
"scope": "project",
"apply_count": 12
}
Key Fields:
trigger: Natural language condition for when to applyaction: Specific, actionable instructionconfidence: 0.3 (new) to 0.95 (max), increases +0.05 per applicationdomain: Category (testing, security, architecture, style, etc.)scope: "project" or "global"apply_count: Successful application count
Six Commands
/instinct-status
Purpose: Display all instincts with filters.
Usage:
/instinct-status- Show all/instinct-status --domain testing- Filter by domain/instinct-status --confidence 0.7- Minimum confidence
Output: Lists PROJECT and GLOBAL instincts with confidence, domain, apply count.
/instinct-export
Purpose: Export instincts to JSON for sharing.
Usage:
/instinct-export- Export to stdout/instinct-export --output instincts.json- Save to file
Format: Portable JSON with project metadata removed.
/instinct-import
Purpose: Import instincts from teammates or other projects.
Usage:
/instinct-import instincts.json- Merge imported instincts
Conflict Resolution: If ID exists, keep higher confidence version.
/evolve
Purpose: Cluster related instincts into skill suggestions.
Algorithm:
- Group by domain
- Find instincts with shared trigger patterns (cosine similarity > 0.7)
- Suggest skill structure with clustered actions
Output: Proposed SKILL.md structure for manual review.
/promote
Purpose: Manually promote project instinct to global.
Requirements:
- Confidence >= 0.8
- Instinct scope = "project"
Effect: Moves instinct to global_instincts array, applies to all projects.
/projects
Purpose: List all projects with learned instincts.
Output: Project names, remote URLs, instinct counts, last updated.
Storage Architecture
Location: ~/.claude/homunculus/projects/<git-remote-hash>/instincts.json
Why git-remote-hash: Unique per repository, consistent across local clones and team members.
Fallback: If no git remote, use directory path hash.
Structure:
{
"project": {
"name": "psc_comet",
"git_remote": "https://github.com/...",
"remote_hash": "a3f5b9c2",
"created": "2026-03-28T20:00:00Z"
},
"instincts": [...], // project-scoped
"global_instincts": [...] // applies to all projects
}
Atomic Writes: Write to temp file, then rename to prevent corruption.
Auto-Promotion
Trigger: Instinct applied in 2+ projects AND confidence >= 0.8.
Process:
- Hook observes instinct applied in project A
- Same instinct (by action similarity) applied in project B
- System detects cross-project usage
- Auto-promotes to global scope
- Logs promotion to learnings.md
Manual Override: User can run /promote inst_NNN to force promotion.
Observation Mechanism
Hook: observe-instinct.sh runs at session Stop (end).
Current: Logs that observation occurred (Phase 8.0.1 stub).
Future Phases:
- Parse tool call traces
- Identify repeated patterns (e.g., always running tests before commit)
- Auto-generate instinct suggestions
- Surface for user approval
Frequency: Stop-only (v0.1.0). Higher frequency optional later (trade-off: interruptions vs learning rate).
Anti-Patterns
Over-promoting: Promoting instincts with confidence < 0.8. Premature globalization spreads unvalidated patterns.
Ignoring low-confidence instincts: Not reviewing instincts with confidence < 0.5. May indicate conflicting patterns or context-specific exceptions.
No evidence review: Accepting instinct suggestions without checking evidence field. Evidence shows when/where pattern was observed.
Manual skill writing instead of /evolve: Writing skills from scratch when instincts exist. /evolve generates skill scaffolds automatically.
No confidence decay: (Not implemented v0.1.0). Future: Confidence should decay if not applied for 30+ days (pattern no longer relevant).
Exporting without review: Sharing instincts.json without removing project-specific secrets or proprietary patterns.
Implementation Notes
Dependencies: Python 3.6+, jsonschema (validation), watchdog (optional, Linux only).
Bootstrap: bash scripts/bootstrap-phase8.sh detects Python, installs deps, validates CLI.
Graceful Degradation: If Python unavailable, hook exits silently. Core psc_comet features unaffected.
Cross-Platform: Tested on Windows (Python 3.14.1) and Linux (Python 3.6+).
Mandatory Checklist
- Verify Python 3.6+ detected or bootstrap instructions provided
- Verify instinct-cli.py implements list, add, apply, promote commands
- Verify storage location at ~/.claude/homunculus/projects//instincts.json
- Verify confidence increases +0.05 per application, capped at 0.95
- Verify promotion requires confidence >= 0.8
- Verify observe-instinct.sh hook registered in settings.json Stop section
- Verify graceful degradation when Python unavailable (hook exits cleanly)
- Verify /instinct-status displays both project and global instincts
- Verify git remote hash used for project identification (fallback to directory hash)
- Verify atomic JSON writes prevent file corruption