Memmy v1.0.4

Highlights

• Added project-scoped desktop workspaces. Open or create a local folder, start conversations inside that project, and keep the Agent's working directory and file access bound to the selected workspace.
• Added built-in browser automation for web and UI tasks, including navigation, snapshots, find, click, typing, console and network inspection, screenshots, viewport controls, and safe local HTML preview.
• Improved long-term Memory by preserving reusable subgoals from complex Agent turns and keeping failure-avoidance lessons isolated from successful strategies.

Agent improvements

• Added the expanded ui-craft Skill for planning, implementing, previewing, and visually validating interfaces with the managed browser workflow.
• Added on-demand Agent Memory onboarding for user-selected local AI tools, including Skill generation, initial history import, validated sync recipes, incremental synchronization, and readiness feedback.
• Browser sessions are isolated per chat, cleaned up after idle time, and wait for Chromium preparation instead of failing during desktop startup.
• Command cancellation and timeouts now clean up descendant process trees instead of leaving commands or output streams running.
• Improved Agent-source synchronization by safely skipping incomplete or oversized turns, preserving retryable incremental watermarks, and separating startup scans from recurring synchronization.

Desktop and account improvements

• Reworked the sidebar and composer around projects, with recent and pinned projects, per-project task groups, standalone tasks, archive views, and clearer operation feedback.
• Added optional invitation codes during sign-in, redemption-result notifications, and referral or promotion reward details supplied by Cloud configuration.
• Added a clearer token-usage breakdown for Agent tasks, Memory summaries, and Memory evolution alongside local BYOK usage.
• Fixed login incorrectly showing an exhausted-token state before current quota information had loaded.
• The Windows Data Management page now displays the actual local Memory storage path and opens that same directory.

Memory and data reliability

• Long, tool-heavy Agent turns can now produce independently searchable Span memories while suppressing duplicate parent and subtask recall results.
• Negative outcomes can become scoped failure-avoidance memories without being promoted into higher-level skills or world knowledge.
• Completed turns become searchable sooner, explicit topic endings preserve the final reply before closing an episode, and chitchat or control-only turns no longer pollute durable Memory.
• Improved processing and retry diagnostics, including clearer exhausted-reasoning errors, preserved failure details, and more precise negative-memory matching.
• Memory database export now creates a consistent SQLite snapshot that includes committed WAL data instead of copying only the primary database file.

Packaging and upgrade notes

• On first start, v1.0.4 runs a one-time workspace migration that binds existing desktop chats to the standalone workspace used by the new project-scoped session model. Existing messages, titles, and pinned or archived state are preserved.
• Packaged desktop builds now include the required native runtime and migration files, with migrations unpacked correctly on Windows and a writable cache used for the local embedding model.
• The Windows local-data path fix identified in the v1.0.3 source is included in the official v1.0.4 installers.
• Expanded best-effort lifecycle and runtime analytics across the desktop app, Agent runtime, Memory CLI, and Agent-source workflows, with China and International edition tagging.

Downloads

• Windows x64 (China): Memmy-1.0.4-win32-x64-cn-signed.exe (github.com)
• Windows x64 (International): Memmy-1.0.4-win32-x64-intl-signed.exe (github.com)
• macOS Apple Silicon (China): Memmy-1.0.4-darwin-arm64-cn-signed.dmg (github.com)
• macOS Apple Silicon (International): Memmy-1.0.4-darwin-arm64-intl-signed.dmg (github.com)

Installation

On Windows, download the appropriate signed .exe and run the installer. On Apple Silicon Macs, download the appropriate signed .dmg, open it, and drag Memmy to Applications.

Checksums

Verify downloads with MD5SUMS.txt (github.com) or SHA256SUMS.txt (github.com) attached to this release.

Full Changelog: https://github.com/MemTensor/memmy-agent/compare/v1.0.3...v1.0.4