evolutiondb-github-sync 0.1.0

Sync GitHub PR / issue activity you're involved in into EvolutionDB long-term memory.

evolutiondb-github-sync

Pull every GitHub PR / issue thread you're involved in — authored, assigned, reviewed, or commented on — into EvolutionDB long-term memory. The MCP server then surfaces them through search_memory, so an AI assistant connected through Claude Desktop / Claude Code can answer "what did we decide on PR #142 last week?" without re-reading the diff.

What gets indexed

For every issue + PR that matches involves:<you> updated:>since:

Record	Key shape
The thread itself	gh_thread_<repo>#<number>
Diff-anchored comments	gh_pr_comment_<comment_id>
Issue-thread comments	gh_issue_comment_<comment_id>
PR reviews with a body	gh_pr_review_<review_id>

Keys are deterministic, so a re-sync UPSERTs instead of duplicating rows. A per-thread watermark on updated_at skips threads that haven't moved since the last pass.

Install

pip install evolutiondb-github-sync

Or from this repo:

pip install -e client/github-sync

Configure

Copy .env.example to .env and fill in:

• GITHUB_TOKEN — classic PAT with repo (or public_repo) + read:user. See https://github.com/settings/tokens.
• GITHUB_USERNAME — the handle to follow.
• EVOSQL_* — point at a running EvolutionDB instance (default localhost:5433 / admin / admin).
• MCP_USER_ID — sticky namespace shared with the MCP server.

Run

# One pass; pull the last 7 days. First-time bootstrap.
evolutiondb-github-sync --once --since 7d

# Daemon — polls every GITHUB_POLL_INTERVAL seconds.
evolutiondb-github-sync --interval 900

# Verify auth + scopes without touching the DB.
evolutiondb-github-sync --once --dry-run

Each pass prints a one-line JSON summary:

{"ok": true, "threads": 12, "comments": 84, "reviews": 6, "errors": 0}

Rate limits

A classic PAT gives you 5,000 REST calls per hour. One sync pass spends roughly 1 search call + (3 calls × open threads), so even 500 active threads fits in under 10% of the budget. The client honours X-RateLimit-Remaining and the secondary-rate-limit Retry-After header — if either fires it sleeps until reset and resumes.

What's not indexed

• Commit messages — git history is already queryable locally.
• Workflow runs / check runs — high volume, low signal for memory recall.
• Notifications — overlaps too much with the issue / comment streams.

Privacy

Records are written to the same EvolutionDB store the MCP server reads from. Anyone with access to that database (and the matching MCP_USER_ID) can recall the indexed content through search_memory. Don't ship a GITHUB_TOKEN that has more scope than you'd hand the AI assistant directly.