Changelog

v3 — Exit Gate + Rich Session Capture (2026-03-22)

Exit Gate (export_session MCP tool) — Agents dump full session context before ending: summary, decisions, bugs, patterns, learnings, files touched, next steps. Creates a rich composite memory + individual typed memories.
Rich session capture — syncmind session end now captures 8 signals instead of 3:
- Git commits (last 2 hours)
- File diffs with names
- Uncommitted changes (staged vs unstaged breakdown)
- Branch name + ahead/behind remote
- Recently modified files (excludes build artifacts)
- Dependency changes from package.json
- Error logs from .next/trace
- Works without git (filesystem-only fallback)
/api/memories/export endpoint — REST API for the Exit Gate
/api/sync GET endpoint — Activity feed via REST (?action=activities)
syncmind hooks command — Installs Exit Gate instructions into CLAUDE.md, .cursor/rules/syncmind.mdc, AGENTS.md, .windsurfrules, plus git hooks and IDE session hooks
--source / -s flag for syncmind session end — Explicitly pass which IDE triggered the session

Better project detection — Falls back through: git remote → git toplevel → cwd basename (3 fallbacks instead of 1)
Better source detection — Checks more env vars including TERM_PROGRAM=Claude
Windows compatibility — All git commands use stdio:['pipe','pipe','pipe'] for proper error handling; PowerShell commands for file discovery
Cross-platform Stop hook — Updated Claude Code Stop hook to pass --source and use robust project detection
Build artifact filtering — Recently modified files excludes .next/, dist/, build/, coverage/, node_modules/

POST /api/memories/export — Session export with structured fields (summary, decisions, bugs, patterns, learnings, files_touched, next_steps)
GET /api/sync?action=activities — Fetch recent activity log entries

Initial release: shared persistent memory for AI coding agents
PowerSync local-first dashboard with real-time sync
Neon Postgres with full-text search + pg_trgm dedup
MCP server with read/write/bump/delete tools
CLI with 9 commands: install, status, write, search, capture, session, hooks, restart, pull
Auto-capture from 13 source types
Freshness scoring, confidence levels, scoped visibility