Conversation Learnings Extractor

Extract structured learnings (lessons, decisions, patterns, dead ends) from exported AI conversations using either a local Ollama model or any OpenAI-compatible API. This skill is designed to work with exports from OpenAI and Anthropic, and pairs well with the chat-history-importer skill for a complete conversation analysis workflow.

Quick Start

Using Ollama (default)

python3 scripts/extract.py --dir /path/to/exports --limit 3 --dry-run
python3 scripts/extract.py --file single-conversation.json
python3 scripts/extract.py --dir /path/to/exports --since 2026-04-01

Using OpenAI-compatible API (e.g., OpenAI, Anthropic Bedrock, etc.)

export OPENAI_API_KEY=sk-...
export OPENAI_BASE_URL=https://api.openai.com/v1  # optional, defaults to OpenAI
python3 scripts/extract.py --dir /path/to/exports --model gpt-4o-mini

How It Works

1. Parse OpenAI/Anthropic JSON exports using bundled parsers (from sibling chat-history-importer skill)
2. Deduplicate via .processed_ids file (skip already-processed chats)
3. Summarize conversation to key excerpts (to fit model context)
4. Extract structured learnings using your chosen model: lessons, decisions, patterns, dead ends
5. Append results to memory/semantic/learnings-from-exports.md

Integration with chat-history-importer

This skill pairs with chat-history-importer:

1. First, run chat-history-importer to ingest raw conversations into episodic memory (memory/episodic/YYYY-MM-DD.md)
2. Then, run this skill to extract structured learnings into semantic memory (memory/semantic/learnings-from-exports.md)

This workflow keeps raw conversation logs separate from actionable insights, enabling better knowledge organization.

Configuration

Using Ollama (Local)

Prerequisites: Ollama running at http://127.0.0.1:11434 (default)

# Use default model (gemma4:26b)
python3 scripts/extract.py --dir /path/to/exports

# Use a different local model
python3 scripts/extract.py --dir /path/to/exports --model llama2

# Custom Ollama endpoint
export OLLAMA_BASE_URL=http://ollama.example.com:11434
python3 scripts/extract.py --dir /path/to/exports

Environment Variables:

• OLLAMA_BASE_URL — Ollama API endpoint (default: http://127.0.0.1:11434)

Using OpenAI-compatible API

Any API supporting the OpenAI /chat/completions endpoint (OpenAI, Bedrock, LM Studio, etc.)

export OPENAI_API_KEY=sk-...
export OPENAI_BASE_URL=https://api.openai.com/v1  # optional
python3 scripts/extract.py --dir /path/to/exports --model gpt-4o-mini

Environment Variables:

• OPENAI_API_KEY — API key (required to enable this mode; if set, OpenAI mode is used instead of Ollama)
• OPENAI_BASE_URL — API base URL (default: https://api.openai.com/v1)

Model auto-selection:

• If OPENAI_API_KEY is set → defaults to gpt-4o-mini
• If OPENAI_API_KEY is not set → defaults to gemma4:26b (Ollama)

Flags

• --dir DIR — Process all JSON files in directory
• --file FILE — Process single file
• --limit N — Process only first N conversations (useful for testing or limiting API costs)
• --since YYYY-MM-DD — Skip conversations before this date
• --model MODEL — Override default model name
• --dry-run — Print output without writing to disk or updating dedup state

Output Format

Results are appended to memory/semantic/learnings-from-exports.md with this structure:

## Chat Title (YYYY-MM-DD)

### Lessons Learned

- [bullet points]

### Decisions Made

- [bullet points]

### Patterns Noticed

- [bullet points]

### Dead Ends

- [bullet points]

Each category is optional — if a conversation doesn't have notable insights for a category, it will show "None".

References

• references/prompt-template.md — The extraction prompt sent to the model
• scripts/extract.py — Main script (reuses parsers from sibling skill)

Implementation Notes

• Tracks processed chat IDs in .processed_ids to avoid re-processing
• Workspace detection: checks OPENCLAW_WORKSPACE env var, falls back to ~/.openclaw/workspace
• Automatically detects OpenAI vs Anthropic export formats
• Truncates long messages for context efficiency