Native support for semantic search and memory rotation (demonstrated in 4-Layer Memory System)

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

When working on multiple projects with Claude Code, developers face:

1. Context loss between sessions - Each new session requires re-explaining project context (5-10 min overhead)
2. Knowledge silos - Solutions discovered in one project aren't accessible in others
3. Manual memory management - No automatic cleanup of outdated information
4. Limited search - Keyword-only search, no semantic understanding

These issues become critical for complex, long-term projects where context accumulation is essential for productivity.

Proposed Solution

Add native support for:

1. Semantic Search in Auto-Memory

• Vector-based search that understands meaning, not just keywords
• Cross-project memory queries
• Multilingual support

1. Memory Rotation API

• Configurable TTL for memory entries (e.g., 24h/14d/permanent)
• Automatic archival of old memories
• Prevents memory bloat while preserving important information

1. Memory Health Commands

• Built-in validation: /memory health
• Detect ghost links, orphaned files, corruption
• Suggest cleanup actions

1. Cross-Project Memory

• Global memory shared across all projects
• Project-specific memory for local context
• Commands like /recall --all to search everywhere

Example config:
{
"memory": {
"rotation": {
"hot": "24h",
"warm": "14d",
"cold": "permanent"
},
"semantic_search": {
"enabled": true,
"cross_project": true
}
}
}

Alternative Solutions

Current approach (community implementation):

I've built a working system that demonstrates these needs: 4-Layer Memory System

• Repository: https://github.com/mergelord/claude-4layer-memory
• License: MIT (open source)

Architecture:

• Layer 4: SEMANTIC (ChromaDB + sentence-transformers)
• Layer 3: COLD (Permanent archive)
• Layer 2: WARM (14 days)
• Layer 1: HOT (24 hours)

Integration:

• SessionStart hook: Load context, validate memory
• Stop hook: Write session summary
• PreCompact hook: Flush semantic DB

Limitations of community approach:

• ❌ Requires external dependencies (ChromaDB, sentence-transformers)
• ❌ ~500MB model download
• ❌ Complex setup for new users
• ❌ Maintenance overhead

Benefits of native support:

• ✅ Zero setup complexity
• ✅ Better integration with Claude Code
• ✅ Consistent UX
• ✅ Lower barrier to entry

Priority

Low - Nice to have

Feature Category

CLI commands and flags

Use Case Example

Real-world results after 3 weeks:*

Projects completed:

• 4-Layer Memory System (~3,000 lines)
• Flight Simulator Integration (~14,878 lines)
• Total: ~18,000 lines in 6 weeks

Productivity improvements:

• Context recovery: 5-10 min → <1 min (90% reduction)
• Overall productivity: 3-4x multiplier
• Code quality: Pylint 9.5+/10, 0 security issues (Bandit)

Example workflow:

1. Cross-Project Learning

/recall --all "how to handle Unicode errors"

Results:
[project-A] decisions.md - UTF-8 enforcement solution
[project-B] feedback.md - Windows console encoding fix

→ Reused solution, saved 2 hours of debugging

1. Automatic Context Recovery

Session starts → loads:

• Last 10 events (HOT layer)
• Important decisions (WARM layer)
• Relevant semantic chunks (L4 layer)

→ Zero manual explanation needed

1. Memory Health

/memory health

Output:
✓ HOT: 3 entries (within 24h)
✓ WARM: 12 entries (within 14d)
✓ L4 index: 489 chunks
✓ No ghost links

→ Automatic validation, no manual maintenance

Community adoption:

• Published on GitHub 2 days ago
• MIT license (open for community use)
• Full documentation, installation scripts
• Cross-platform (Windows/Linux/macOS)

Technical stats:

• Memory validation: <0.5s (quick mode)
• Semantic search: 1-3s
• Memory footprint: ~50MB
• Disk usage: ~500MB (including embeddings model)

Additional Context

Questions for Anthropic team:

1. Are there plans for native semantic search in Claude Code?
2. Would you consider adding memory rotation APIs?
3. Is there interest in featuring community projects like this in official docs?

Community value:

• Demonstrates advanced hook usage patterns
• Shows real-world demand for these features
• Provides reference implementation
• Can inform future Claude Code development

Credits:
This implementation integrates ideas from:

• qwwiwi's 4-layer architecture concept
• cablate's atomic wiki system
• Custom L4 semantic layer, auto-discovery, multilingual support

Repository: https://github.com/mergelord/claude-4layer-memory
Status: Production-ready, actively maintained
License: MIT