A new memory system, the Easy rebrand, a built-in sandboxed coding agent and a major flow compile pipeline rebuild, alongside Contacts becoming first-class and a redesigned chat composer and model picker.
New
Seven memory styles
Personal Notebook, Wiki, Support KB, Agent History, Document Archive, Coding Agent and Entity Profile - each tuned for a different kind of remembering.
New
Memories page
A new Memories page in the sidebar with Files, Search and Trash tabs so you can browse what an agent has remembered. Search is wired into Fold and robust to renames.
New
Engram sleep agent
A dedicated Engram sleep agent runs in the background, filing unsorted notes from the inbox into the right concept folders and keeping memory tidy.
Improvement
Smarter saves
Memories route to the right place via a registry rather than dumping into one bucket, with agents discovering existing slugs instead of guessing names.
Improvement
Memory is opt-in
The Forge no longer attaches a default memory system. Easy nudges you to set one up after creating an agent if it makes sense.
New
Memory Building skill
A guided KB capture skill lets you sit with an agent, dump what it should know and end up with a properly organised knowledge base.
Improvement
Existing workspaces upgrade
Older engram folders get pulled into the new layout cleanly on first run.
New
Hello, Easy
The platform formerly known as Agency Agent and SiGMA is now Easy, with a new Manrope ExtraBold wordmark across the UI and a fresh BrandMark, Layout, Site Settings, icon generator and Figma tokens.
Improvement
Easy rebrand across the repo
End-to-end rename across the repo, workspaces and code-workspace metadata, with stale agency-agent strings scrubbed from templates, skills and docs.
New
Agents that talk
Agents can now reply with audio inline, right in the chat thread, with a compact native player for short clips and a richer waveform view for longer ones.
Improvement
Voice autoplay and captions
Short voice messages autoplay once with white captions overlaid, and transcripts sit alongside every clip.
Improvement
One clip at a time
Only one audio clip plays at a time so two agents do not talk over each other, with persistent caching so clips do not reload on every conversation update.
New
Claw Code
A new sandboxed coding agent is built into the platform with skills, tooling, container image and setup docs, with coordinator skills routing coding tasks its way automatically.
New
Live Translator
A new in-chat Live Translator skill is available, useful for multilingual customer chats or international teammates.
New
Webhooks (at last)
Flows can now receive webhooks, not just send them, with persistent webhook triggers letting a flow wait for an external service to call it.
New
Webhooks page
A new manage/webhooks page shows every webhook across flow triggers, callbacks and event topics in one unified list.
Improvement
Webhooks preferred over polling
The compile pipeline now prefers webhooks over polling wherever an external service supports them, with webhook registration happening as part of the build.
Improvement
Clean public webhook URLs
Public webhook URLs follow a clean, predictable pattern.
New
Contacts first-class
Contacts are now a proper first-class concept and lists are the partition - newsletter subscribers, paying customers, beta testers - not workspaces.
Improvement
Contact UI rebuild
A card grid layout instead of a table, with a unified editor, drag-to-reorder priority and batch paste for dropping a block of contacts cleanly.
Improvement
Flow builders contact-aware
Flow builders are aware of the contacts system and can wire it in automatically, treating people as first-class contacts everywhere.
New
Three-session flow pipeline
Planning, architecture and compiling now run as separate specialists handing off to each other instead of one agent trying to do all three.
New
Real entry points for flows
Flows can declare themselves as directly callable so you or another flow can just invoke them with no trigger required.
New
Email as a flow channel
A new email channel adapter contract means flows can ask questions and receive answers over email cleanly with proper subject, from and auth wiring.
Improvement
No more mediator subagent
The old mediator subagent pattern for asks is gone. Asks go directly via the channel now.
Improvement
Strict flow input validation
Flows refuse to run on bad input rather than fail halfway through, and the main session waits properly for completion.
Improvement
Coordinator skill split
The coordinator skill is split into a router plus eight focused sub-skills, with less prompt bloat and faster compiles for simple jobs.
New
Channels as skills
Each channel - Telegram, Voice, Webchat, Email - now ships its own auto-loaded skill, so agents get the right manners loaded automatically.
Improvement
Faster, tidier chat
Settled messages stop re-rendering on every streaming chunk and markdown rendering is memoised, so old messages are not re-parsed.
Improvement
No half-tag flashes
Mid-stream audio, video and iframe tags are hidden until they complete.
Improvement
Stable audio mounts
Audio elements get stable mounts and do not reload mid-conversation.
Improvement
Better rewind behaviour
Rewinding cleanly drops the message you rewound to instead of orphaning it.
Improvement
Files are clickable links
Workspace files in chat are always proper clickable links, never a bare filename.
Improvement
Cleaner LLM error chips
A cleaner LLM error chip with a classifier tells you what kind of error it was rather than a wall of red.
New
Chat composer redesign
A fully-pill composer with a single-row OpenAI-style layout and a cascading model menu with real brand glyphs.
Improvement
Draft and save model selection
Pick a model and confirm instead of switching on every click. Vision attachments no longer silently flip your model.
New
New Dashboard
A new Dashboard landing page with Recent Jobs, Recent Messages and an Agent Row showing your team at a glance.
Improvement
Tighter job coordination
Dedicated job-planning skill, an operator engagement gate, faster wake-ups, atomic task spawning and a per-run Team tab.
Improvement
Stuck-run detection
Stuck-run detection automatically spots flows that have hung, and restarts preserve worker sessions instead of orphaning them.
Improvement
Automations built in
The Automations page is now first-class - loads same-origin in production, auto-starts on visit and inherits your Easy branding.
Improvement
Better Forge handoff
Forge stays on page after handoff and confirmation happens in your current chat, not as an out-of-band server commit.
Improvement
Smarter notification bell
The bell rings and flashes on task completion or failure, shows a task button when something needs attention, and Mark all read clears pending operator requests.
Improvement
Memories page polish
The Memories page was redesigned to match the Figma spec, with polish on the detail view.
Improvement
Status palette
Across Contacts and Webhooks, alarming red error banners are replaced with friendly status-warn cards explaining why and what to do.
Improvement
Path-segment routing
Job tabs have real URLs and the Team tab is deep-linkable.
Improvement
Reliability under load
WAL mode on the orchestration database with a 5-second busy timeout means far fewer database-locked hiccups under load.
Improvement
Bigger tool results
Tool results can be 6x bigger (8K to 50K characters) before being truncated.
Improvement
OpenAI Realtime STT
OpenAI Realtime STT models now appear in the voice picker, with automatic migration of older batch IDs.
Fix
Email channel fixes
Four stacked inbound-email bugs were tracked down and fixed - reply threading, recipient resolution and content extraction all work cleanly now.
Fix
WhatsApp connect
WhatsApp connect no longer crashes on a schema mismatch.
Fix
Chat auth errors
Chat auth errors show actionable toast text instead of opaque upstream messages.
Fix
Handoff retry
Handoff retries once on a transient session-not-found to bridge a file-flush race.
Fix
Skills load cap
Skills load is capped at 3 per call with a preview mode so you cannot flood your prompt.
Fix
Spawn no-op guard
Empty or no-op tasks are blocked before they create useless sessions.