Codecompanion Memory

CodeCompanion Memory Skill

This skill provides access to a searchable history of previous CodeCompanion conversations stored in a ChromaDB vector database. It uses semantic search to find relevant past conversations based on natural language queries.

When to Use This Skill

Invoke this skill when:

• User references past conversations: Phrases like "we discussed this before", "remember when", "last time we talked about"
• Recurring problems: The user encounters an issue that might have been solved previously
• Context would be helpful: Previous solutions, decisions, or architectural discussions could inform the current task
• User asks about their codebase: Questions about past work, project history, or previous implementations
• Building on previous work: Extending or modifying solutions from past conversations

Usage

~/.claude/skills/codecompanion-memory/query.sh --query "your search query"

Parameters

• --query TEXT or -q TEXT: The search query (required)

  • Use natural language
  • Be specific but not too narrow
  • Examples: "fixing Docker networking", "Go error handling patterns", "setting up ChromaDB"

• --count NUMBER or -n NUMBER: Number of results to return (default: 5)

  • More results = more context but also more noise
  • Recommended: 3-5 for focused queries, 10+ for exploratory searches

• --verbose or -v: Show full document content

  • Without this flag, only previews (first 200 chars) are shown
  • Use when you need to read the full conversation summary

• --project PATH or -p PATH: Override default project root (default: ~/codecompanion-history/summaries)

• --help or -h: Show help message

Example Queries

Finding Past Solutions

# User: "I'm getting a CORS error, have we dealt with this before?"
~/.claude/skills/codecompanion-memory/query.sh \
  --query "CORS error HTTP request" \
  --count 3

Architectural Decisions

# User: "What database did we choose for the user service?"
~/.claude/skills/codecompanion-memory/query.sh \
  --query "database choice user service" \
  --count 5 \
  --verbose

Code Patterns

# User: "How did I implement authentication last time?"
~/.claude/skills/codecompanion-memory/query.sh \
  --query "authentication implementation pattern" \
  --count 3 \
  --verbose

Output Format

Results are displayed in order of relevance:

Searching in: /home/djipey/codecompanion-history/summaries
Query: Arduino

--- Result 1 ---
Path: 1759511459.md
Preview: ## Code Context
**Files Modified**: ardoise.ino (main focus: refactoring BLE write logic)...

--- Result 2 ---
Path: 1759512219.md
Preview: ## Code Context
**Files Modified**: ardoise.ino (main Arduino sketch)...

Each result includes:

• Path: The markdown file containing the full conversation summary
• Preview/Document: Either a short preview or full content (with --verbose)

You can read the full conversation summary files at: ~/codecompanion-history/summaries/<chat_id>.md

How It Works

User Query
    ↓
Bash Script (query.sh)
    ↓
VectorCode CLI
    ↓
ChromaDB (local instance)
    ↓
Semantic Search Results

The skill:

1. Wraps the VectorCode CLI tool
2. Queries a local ChromaDB database
3. Uses semantic embeddings (SentenceTransformer)
4. Returns results sorted by relevance

Database Information

• Database Location: ~/.local/share/vectorcode/chromadb/chroma.sqlite3
• Summaries: ~/codecompanion-history/summaries/*.md
• Embedding Model: SentenceTransformer (all-MiniLM-L6-v2)

The database uses semantic embeddings, meaning:

• Queries find conceptually similar content, not just keyword matches
• Synonyms and related terms are automatically understood
• Context and meaning are preserved across different phrasings

Installation

The skill is installed at ~/.claude/skills/codecompanion-memory/.

If you need to reinstall or update:

mkdir -p ~/.claude/skills/codecompanion-memory

# Copy required files from source
cp /path/to/source/SKILL.md \
   ~/.claude/skills/codecompanion-memory/
cp /path/to/source/query.sh \
   ~/.claude/skills/codecompanion-memory/

# Make script executable
chmod +x ~/.claude/skills/codecompanion-memory/query.sh

Dependencies

• bash: Shell interpreter
• python3: For JSON parsing in the script
• vectorcode: CLI tool for querying ChromaDB

Installing VectorCode

If VectorCode is not installed:

# Using pipx (recommended)
pipx install vectorcode

# Or using pip
pip install --user vectorcode

# Verify installation
which vectorcode

Performance

• First query: ~3s (VectorCode cold start)
• Subsequent queries: ~1-2s
• Memory usage: ~150MB (VectorCode + ChromaDB server)
• Disk space: ~3KB (just the script)

Troubleshooting

"vectorcode command not found"

# Install VectorCode
pipx install vectorcode

# Or check installation
which vectorcode

"No results found"

• Try broader queries: "Docker" instead of "Docker Compose networking with custom bridge"
• Check if conversations are indexed: ls ~/codecompanion-history/summaries/
• Verify database exists: ls ~/.local/share/vectorcode/chromadb/

VectorCode configuration

If you have a ChromaDB Docker container and want VectorCode to use it, create:

# ~/codecompanion-history/summaries/.vectorcode.toml
[default]
db_url = "http://localhost:8001"
db_path = ""

Note: The skill works fine with VectorCode's default local database.

Best Practices

1. Start broad, then narrow: Begin with general queries, then refine based on results
2. Use verbose mode selectively: Only when you need full context to answer the user's question
3. Combine with code search: This skill finds conversations; use code search tools to find actual implementations
4. Cite your sources: Tell the user which conversation(s) you found the information in
5. Verify information: Past solutions might be outdated; always validate before applying

Example Workflow

User: "I'm stuck on the same authentication bug we had last month"

Step 1: Search for relevant conversations
→ ./query.sh --query "authentication bug fix" --count 5

Step 2: Review results, identify the most relevant conversation
→ Result 3 seems most relevant (Path: 1763841695.md)

Step 3: Read full summary if needed
→ cat ~/codecompanion-history/summaries/1763841695.md

Step 4: Apply the solution or adapt it to current context

Step 5: Inform user
→ "I found a similar issue we solved in conversation 1763841695. 
   The problem was related to session token expiration. 
   Here's what we did..."

Technical Details

Script Implementation

The skill is a simple Bash wrapper that:

1. Validates and parses command-line arguments
2. Calls vectorcode query with appropriate parameters
3. Formats the JSON output for display
4. Handles errors gracefully

Why VectorCode?

VectorCode handles:

• ChromaDB connection and configuration
• Embedding generation for queries
• Collection management
• Query execution

This makes the skill implementation simple and reliable.

Integration Notes

This skill is designed to work seamlessly with Claude Code's skill system:

• Automatically loaded when referenced or when the user mentions past conversations
• Runs independently without blocking other operations
• Returns structured data that can be easily parsed and presented to the user

Status: ✅ Production ready
Location: ~/.claude/skills/codecompanion-memory/
Last Updated: 2025-12-15