eliasoulkadi chromadb · Memory Storage
ChromaDB is the vector database that powers Shokunin's persistent AI memory. It stores embeddings for semantic search across sessions. Built on SQLite with an ONNX embedding model for local-first, offline-capable operation.
Workflow
Step 1: Check storage status
python ~/.shokunin/scripts/chroma-helper.py count
Output: total entries, collection count, disk size on disk, last write timestamp.
python ~/.shokunin/scripts/chroma-helper.py stats
Output: entries per type (decision, file, command, checkpoint, session_end), per project, per date range.
Step 2: Search stored entries
python ~/.shokunin/scripts/chroma-helper.py search "query text" "project-name" 10
Searches via vector similarity + BM25 keyword hybrid. Results sorted by relevance with metadata attached.
python ~/.shokunin/scripts/chroma-helper.py search "error handling" "" 20 --type decision
Filter by entry type. Leave project empty to search all projects.
Step 3: List recent entries
python ~/.shokunin/scripts/chroma-helper.py recent 10
Shows the 10 most recently inserted entries across all projects. Use for quick audit.
python ~/.shokunin/scripts/chroma-helper.py recent 50 --project myproject
Filter recent entries to a specific project.
Step 4: Backup the database
$backupDir = "$env:USERPROFILE\.shokunin\backups\chroma-$(Get-Date -Format 'yyyyMMdd-HHmmss')"
New-Item -ItemType Directory -Path $backupDir -Force | Out-Null
Compress-Archive -Path "$env:USERPROFILE\.shokunin\memory\chroma_db" -DestinationPath "$backupDir\chroma_db.zip" -Force
Also back up session logs:
Compress-Archive -Path "$env:USERPROFILE\.shokunin\memory\sessions" -DestinationPath "$backupDir\sessions.zip" -Force
Verify backup integrity:
Add-Type -AssemblyName System.IO.Compression.FileSystem
[System.IO.Compression.ZipFile]::OpenRead("$backupDir\chroma_db.zip").Entries.Count
Step 5: Delete entries by session prefix
python ~/.shokunin/scripts/chroma-helper.py delete "session-prefix"
Deletes all entries whose session_id starts with the given prefix. Non-reversible. Backup first.
python ~/.shokunin/scripts/chroma-helper.py delete --project "old-project"
Delete all entries for a project. Useful when archiving finished projects.
Step 6: Full reset (requires confirmation)
Remove-Item -Recurse -Force "$env:USERPROFILE\.shokunin\memory\chroma_db"
# Restart MCP server to recreate the database
The database is auto-created on next MCP server startup. All sessions, embeddings, and metadata are permanently lost. Only do this when migrating or recovering from corruption.
Storage Details
| Location | Purpose |
|---|---|
~/.shokunin/memory/chroma_db/ |
ChromaDB persistent files (SQLite + embeddings) |
~/.shokunin/memory/sessions/ |
Per-session JSONL logs and Markdown summaries |
~/.shokunin/memory/mcp-server.log |
MCP server log file (auto-rotated at 5MB) |
~/.shokunin/backups/ |
Timestamped backup archives |
~/.shokunin/scripts/chroma-helper.py |
CLI management script |
Data Model
| Concept | Description |
|---|---|
| Collection | ChromaDB namespace for embeddings (one per project by default) |
| Embedding | Vector representation of text for semantic similarity search |
| Metadata | Key-value pairs attached to each entry (type, tags, project, session_id, timestamp) |
| Document | The raw text content stored alongside the embedding |
| BM25 Index | Sparse retrieval index that complements vector search (keyword matching) |
| Result Fusion | Merges vector + BM25 results for higher recall (reciprocal rank fusion) |
| Entry Types | decision, file, command, preference, checkpoint, session_end, general |
Session Management
| Command | Description |
|---|---|
python ~/.shokunin/scripts/chroma-helper.py session list [N] |
List recent sessions with metadata |
python ~/.shokunin/scripts/chroma-helper.py session continue <id> |
Load full session context |
python ~/.shokunin/scripts/chroma-helper.py session summary <id> |
Show session summary |
python ~/.shokunin/scripts/chroma-helper.py save "<text>" "<id>" "<type>" "<tags>" "<project>" |
Save entry |
Error Handling
| Error | Cause | Fix |
|---|---|---|
| ChromaDB not found | Package not installed | pip install chromadb |
| Collection empty | No data stored yet | Normal on first use. Start storing context. |
| Slow first query | Downloading ONNX model | First query downloads ~79MB. Subsequent queries are instant. |
| Permission denied | File locked by another process | Close other MCP server instances |
| Corrupted database | Unexpected shutdown or disk full | Restore from backup or delete and let the system reinitialize |
| Memory usage high | Large collection over time | Run consolidate_memories to merge old entries |
| Embedding dimension mismatch | Model was upgraded or changed | Delete chroma_db/ and reindex from session logs |
| UUID collision | Race condition in parallel writes | ChromaDB handles this internally; retry the insert |
| SQLite disk I/O error | Filesystem full or read-only | Free disk space (need ~2x DB size); check permissions |
Anti-Patterns
| Pattern | Problem | Fix |
|---|---|---|
Deleting chroma_db/ without backup |
Permanent data loss | Always backup first (Step 4) |
| Running multiple MCP servers | Database locked, data corruption | Only one MCP server instance at a time |
| Manual SQL manipulation | Breaks ChromaDB invariants | Use chroma-helper.py CLI or MCP tools only |
| Ignoring session_list filters | Seeing test/healthcheck entries | session_list automatically filters noise |
| Storing extremely large text | Bloating vector index | ChromaDB truncates at 500 chars |
| Never consolidating | Database grows indefinitely | Run consolidate_memories periodically |
| Skipping backups | No recovery from disk failure | Weekly backup is configured via Task Scheduler |
| Using raw SQLite instead of ChromaDB API | Schema and index corruption | Always go through ChromaDB's Python client |
| Searching without project filter for broad queries | Too many irrelevant results | Always filter by project when context is known |
Automation
Weekly backup every Sunday via Windows Task Scheduler (pre-configured). Manual backup via Step 4 above. Logs auto-rotate at 5MB to prevent unbounded growth. The consolidate_memories function runs automatically when entry count exceeds 10,000 per project.
Recovery Procedures
Restore from backup
$backup = Get-ChildItem "$env:USERPROFILE\.shokunin\backups" -Directory | Sort-Object LastWriteTime -Descending | Select-Object -First 1
Expand-Archive -Path "$backup\chroma_db.zip" -DestinationPath "$env:USERPROFILE\.shokunin\memory\chroma_db" -Force
Rebuild from session logs
If backups are unavailable, session JSONL logs contain all raw text. Replay them:
python ~/.shokunin/scripts/chroma-helper.py rebuild --from-sessions
Verify database health
python ~/.shokunin/scripts/chroma-helper.py verify
Checks: SQLite integrity, embedding count vs document count, orphan metadata, collection schema version.
Checklist
- ChromaDB client initialized lazily (not at import time)
- Collection name sanitized (alphanumeric + hyphens only)
- Embedding dimension matches between writes and queries
- Backups configured before destructive operations (delete/reset)
- Search results validated — empty results handled gracefully
Sources
- ChromaDB documentation (docs.trychroma.com)
- OpenAI text-embedding-ada-002 (or local ONNX embedder)
- SQLite documentation (sqlite.org)
- DuckDB (ChromaDB's internal query engine)
- MCP Protocol (modelcontextprotocol.io)
- Shokunin Memory System Architecture (ARCHITECTURE.md)