v1.4.0
v1.4.0 — File Sync, AI Instruction Rewrite & Space Switcher
Two milestones ship in this release: Anchor Memories (v1.4) and AI Instruction & Search Quality (v1.5).
File Sync (Anchor Memories)
Covalence can now watch folders on disk and auto-index their contents as searchable memories.
- Watched folders — add any directory from Settings → Files. Covalence indexes markdown, text, and config files using FSEvents for live updates.
- Smart defaults — ships with sensible ignore patterns (
.git/,node_modules/,.env*, etc.) and a 1 MB file size cap. Customize with a freeform gitignore-style editor. - Source-tier ranking — search results now rank Core memories first, then MCP/App, then File memories. The
fileTierWeightconstant (0.20) can be tuned via the benchmark tool. - YAML frontmatter — parsed from markdown files and stored as structured metadata.
- Space assignment — each watched folder is bound to a space. Move folders between spaces from the context menu.
- Keep or delete — when removing a watched folder, choose whether to keep its memories as orphans or delete them.
AI Instruction Rewrite
The behavioral instruction that tells AI clients how to use Covalence has been completely rewritten.
- Two-axis taxonomy — memories are now categorized by source (MCP, App, File) and intent (Core flag). The instruction guides AI clients on when to use
source: "file"filters. - MCP server-level instructions — Covalence now delivers its behavioral guidance automatically at MCP connect time via
Server.init(instructions:). Compliant clients (Claude Code, Claude Desktop) pick it up without any manual configuration. - Windsurf support — added as the 4th onboarding tab with the correct MCP config path.
Space Switcher
- Filter by space — a compact picker in the main window header lets you scope the browse list and search results to a single space. "All Spaces" is the default. Selection persists across restarts.
Retrieval Quality Benchmark
- cov-benchmark — a new CLI tool that runs 10 queries across multiple categories against the live database and scores retrieval quality. Useful for tuning ranking parameters. Current score: 30/30.
Upgrading
Sparkle auto-update will install v1.4.0 automatically.