chat-learnings-extractor

Extract structured learnings (lessons, decisions, patterns, dead ends) from AI conversation exports using a local Ollama model or any OpenAI-compatible API. Pairs with chat-history-importer. Trigger phrases: extract learnings from conversations, analyze chat exports, mine conversation insights, extract lessons from chats, chat learnings extractor.

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.

Data Handling and Safety

This skill reads private conversation exports and writes model-generated output into your agent's persistent memory. Treat both directions as sensitive.

• Dry run first. Always start with --dry-run and read the output before letting it write anything to memory.
• Prefer local processing. The default Ollama path runs against http://127.0.0.1:11434 on loopback, so export contents never leave your machine.
• Remote endpoints send your exports. When OPENAI_API_KEY is set, conversation content and the API key are sent to the configured endpoint. Only set OPENAI_BASE_URL to a trusted HTTPS provider you control.
• Review before you trust. Model output is unreviewed text written into persistent agent memory. Read the generated learnings before relying on them.
• Scope the input. Narrow --dir and --since so you import only what you intend.

See references/model-modes-and-categories.md for the OpenAI-compatible API setup and the full extraction category list.

Configuration, model modes, and the full output format are documented in references/model-modes-and-categories.md.

Quick Start

Using Ollama (default)

bash

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

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.

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

References

• references/model-modes-and-categories.md — Configuration, model modes, output format
• 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