v0.2.0

Added

• Agents read the channel. Every turn's prompt is rebuilt from what is stored: the channel's
  running summary, the five latest messages from every person and agent, and up to three of the
  agent's own memories that match what was said. Nothing is lost on a restart, and every agent
  sees the others' replies.
• Channel summaries. Once about 6,000 characters of chat pile up, the router's small model
  rewrites the channel's summary in the background: decisions, open questions, owners, figures and
  live objects, each citing the message id it came from.
• search_room and read_chat. Keyword search over a channel's whole message history and the
  text of its docs, sheets, slides, canvases and uploaded PDF, .xlsx and text files, from a
  derived Postgres index kept current after every save; and paging through the chat around or
  before any message.
• Scripts call the room's tools. A run_code script imports room (Python) or room.mjs
  (Node) and calls the agent's own tools by name. Calls go over a socket in the jail to the
  runner's control server and back to AirLink, which runs them as the agent that wrote the code,
  at most 50 a run; the jail's network fence is unchanged. A scheduled refresh calls them again,
  so a dashboard pulls live room data.
• Reply mode per channel. Auto, or only when an agent is @-mentioned (or named, in a huddle).
• Huddle voices from Cartesia and Deepgram as well as ElevenLabs and OpenAI; a huddle stays
  shut until agents have a voice.
• Onboarding Home with setup steps and starter prompts, and drafted agent instructions in
  Settings.
• live-server: an onStored hook after each debounced save, and browserHistoryFrames to keep
  fewer (or no) screenshot keyframes.

Changed

• In Auto, chat routing leaves a message to people only when it is plainly addressed to another
  person; small talk and questions no role covers go to the closest agent.
• A runner reaches the channel's desktop only once it has a dashboard or a refresh schedule.
• AirLink keeps no browser or desktop screenshot history; nothing read it.

@memo-air/agent

Patch Changes

• Updated dependencies []:
  • @memo-air/protocol@0.2.0
  • @memo-air/decide@0.2.0
  • @memo-air/space@0.2.0

@memo-air/agents

Minor Changes

• Agents read the channel and search the room: each turn carries the channel's running summary, the five latest messages and the agent's matching memories; search_room and read_chat reach everything older. Scripts run by run_code can call the room's tools. Channels get a reply mode (Auto or on mention). Huddle voices add Cartesia and Deepgram. live-server gains an onStored hook and a browserHistoryFrames option.

Patch Changes

• Updated dependencies []:
  • @memo-air/decide@0.2.0

@memo-air/bridge

Minor Changes

• Agents read the channel and search the room: each turn carries the channel's running summary, the five latest messages and the agent's matching memories; search_room and read_chat reach everything older. Scripts run by run_code can call the room's tools. Channels get a reply mode (Auto or on mention). Huddle voices add Cartesia and Deepgram. live-server gains an onStored hook and a browserHistoryFrames option.

Patch Changes

• Updated dependencies []:
  • @memo-air/agents@0.2.0
  • @memo-air/protocol@0.2.0
  • @memo-air/agent@0.2.0
  • @memo-air/tools@0.2.0
  • @memo-air/decide@0.2.0
  • @memo-air/memory@0.2.0

@memo-air/browser

No changes in this release.

@memo-air/chat

Minor Changes

• Agents read the channel and search the room: each turn carries the channel's running summary, the five latest messages and the agent's matching memories; search_room and read_chat reach everything older. Scripts run by run_code can call the room's tools. Channels get a reply mode (Auto or on mention). Huddle voices add Cartesia and Deepgram. live-server gains an onStored hook and a browserHistoryFrames option.

@memo-air/client

Patch Changes

• Updated dependencies []:
  • @memo-air/protocol@0.2.0

@memo-air/react

Minor Changes

• Agents read the channel and search the room: each turn carries the channel's running summary, the five latest messages and the agent's matching memories; search_room and read_chat reach everything older. Scripts run by run_code can call the room's tools. Channels get a reply mode (Auto or on mention). Huddle voices add Cartesia and Deepgram. live-server gains an onStored hook and a browserHistoryFrames option.

Patch Changes

• Updated dependencies []:
  • @memo-air/protocol@0.2.0
  • @memo-air/browser@0.2.0
  • @memo-air/canvas@0.2.0
  • @memo-air/desktop@0.2.0
  • @memo-air/media@0.2.0
  • @memo-air/richtext@0.2.0
  • @memo-air/runner@0.2.0
  • @memo-air/sheet@0.2.0
  • @memo-air/slides@0.2.0
  • @memo-air/space@0.2.0

@memo-air/server

Minor Changes

• Agents read the channel and search the room: each turn carries the channel's running summary, the five latest messages and the agent's matching memories; search_room and read_chat reach everything older. Scripts run by run_code can call the room's tools. Channels get a reply mode (Auto or on mention). Huddle voices add Cartesia and Deepgram. live-server gains an onStored hook and a browserHistoryFrames option.

Patch Changes

• Updated dependencies []:
  • @memo-air/agents@0.2.0
  • @memo-air/chat@0.2.0
  • @memo-air/protocol@0.2.0
  • @memo-air/browser@0.2.0
  • @memo-air/canvas@0.2.0
  • @memo-air/decide@0.2.0
  • @memo-air/desktop@0.2.0
  • @memo-air/media@0.2.0
  • @memo-air/richtext@0.2.0
  • @memo-air/sheet@0.2.0
  • @memo-air/slides@0.2.0
  • @memo-air/space@0.2.0

@memo-air/slides

Patch Changes

• Updated dependencies []:
  • @memo-air/richtext@0.2.0

@memo-air/tools

Patch Changes

• Updated dependencies []:
  • @memo-air/protocol@0.2.0
  • @memo-air/agent@0.2.0
  • @memo-air/sheet@0.2.0

@memo-air/voice

Minor Changes

• Agents read the channel and search the room: each turn carries the channel's running summary, the five latest messages and the agent's matching memories; search_room and read_chat reach everything older. Scripts run by run_code can call the room's tools. Channels get a reply mode (Auto or on mention). Huddle voices add Cartesia and Deepgram. live-server gains an onStored hook and a browserHistoryFrames option.

Patch Changes

• Updated dependencies []:
  • @memo-air/agents@0.2.0

v0.1.0

Added

• Channels. #general open to everyone; every other channel invite-only, with members (people
  and agents), join requests and a members dialog. Names are the identity: no accounts, a signed
  name cookie.
• Chat. Messages with mentions, attachments, reactions and Markdown; agents stream their
  replies; a typing line.
• The space. Each channel's objects as windows on a shared desktop, with a dock; open, close to
  an icon, minimize, maximize, for everyone; a bin for removed objects.
• Live objects. Rich-text docs with comments, suggestions, history and Markdown export; sheets
  with formulas and .xlsx export; slides with themes and .pptx export; an Excalidraw canvas;
  uploaded files up to 25 MB (PDF, .xlsx and text readable by agents).
• Shared browser. neko's Chromium over WebRTC, driven by people and agents, read by agents as
  an accessibility outline, with every page connection through a pinning proxy that refuses
  private addresses.
• Desktop. A Linux desktop in Docker (XFCE, terminal, editor, files; LibreOffice and Firefox
  with DESKTOP_EXTRAS=1) behind a token-gated control server; E2B as an optional provider.
• Huddle. Voice and video over LiveKit. Full-duplex agent voice: server-side transcription per
  person, on-device turn-end detection, one speaking floor, barge-in with cut marks, an echo filter,
  a distinct voice and speed per agent, and a talker that hands long work to the agent's harness.
  Without a speech key, per-person transcription in the browser and agents answering in chat.
  "Share signals" sends present/away and gestures computed in the browser.
• Agents. Flat agents with a role, instructions, tags, access, harness, model, memory folder
  and voice; agent pages in Settings. Pi in process on keys for Anthropic, OpenAI, Google,
  OpenRouter, Groq, xAI, Mistral or DeepSeek; Claude Code, Codex, Gemini CLI and opencode over ACP
  with the room's tools served as loopback MCP and their own tools refused. Agent cursors.
• Routing and the decision node. Mention, then tag, then the decision node by role, then a
  fallback; room events judged for an unprompted turn; every decision logged with its odds. A small
  model on your key by default; any System One provider (Laya in a compose profile, or Jev) in
  Settings.
• Memory. One OKF folder of Markdown per agent, with recall_memory and remember, gated by
  the decision node; point it at an Obsidian vault.
• live-server. Yjs over Hocuspocus with an append-only Postgres update log, an event log, agent
  ops over HTTP under per-agent, per-doc tokens, presence, /metrics and /healthz; runnable on
  its own (apps/live-server).
• Benchmarks. apps/bench: keystroke, awareness, agent-op, cold-start, concurrency and
  shared-browser latency, with the measured table in the README.
• Project. MIT licence in every package, a dependency licence check, CI (type-check, Biome,
  tests, licences, bench smoke, image build), contributing guide, security policy, documentation
  under docs/.

Coming

• A code runner object: agents run Python and Node in a sandbox and show output and live
  dashboards in the space.
• Published npm packages, npm create airlink, a docs site and an API reference.
• A measured voice latency row.