claude-history-ingest mines the conversation history Claude Code leaves under ~/.claude, plus the desktop app’s local agent sessions, and compiles the useful parts into an Obsidian wiki.
It starts with the memory files Claude already wrote, then reads new conversations, groups what it finds by topic and writes pages with summaries, confidence fields and provenance markers. A manifest records every source file, so later runs only handle new or changed sessions.
What You Get
- Wiki pages for each project, plus global concept, skill, entity and synthesis pages distilled from your sessions
- A manifest that records every ingested source file, so later runs only process the delta
- Provenance markers (extracted, inferred, ambiguous) and a summary line on every page
- A coverage report that says what was sampled and what was skipped
Common Use Cases
- Capture what you learned across months of Claude Code sessions into one searchable vault
- Pull the memory files of every project into a single wiki
- Sync only the sessions you ran since the last ingest
- Build debugging playbooks from the commands recorded in desktop audit logs
What Makes It Unique
It is built around Claude’s own data layout: memory files with their user, feedback, project and reference types, session metadata, desktop audit logs and conversation JSONL. It ranks those sources by value, honors a skip list for projects you want left out, and uses a pre-extraction helper that shrinks transcripts a great deal before reading them.
Under The Hood
The skill is a long set of instructions for the agent. It tells the agent to read config and the vault manifest, list and read files under the Claude history folders, run a Python extraction script and a manifest helper, and call the obsidian-wiki command line to sync the index, log and hot cache. It writes markdown pages and the manifest into your vault. If a QMD search collection is configured it also runs the qmd command to refresh the index. It makes no web requests of its own.