Albert v0.2.1

Albert can now connect to your own kaeru memory over MCP. Keep memory embedded
as before, or let a separate kaeru server hold it while Albert recalls, learns and
reflects through the same memory tools.

---

Bring your own memory

• HTTP or stdio. Connect to a kaeru MCP endpoint over Streamable HTTP, or use
  the local kaeru-mcp --stdio bridge.
• Token authentication. A protected endpoint takes a bearer token from an
  environment variable; the config stores only its name.
• Ready on first startup. If the connected memory has no albert initiative,
  Albert creates it through an ordinary memory record. Later starts leave an
  existing initiative unchanged.
• One memory interface. Recall, capture and automatic reflection use the
  selected backend. MCP memory is a dedicated connection, separate from the
  world-facing connector catalog.

Connection failures stay visible

• Connection and tool calls have time limits. A disconnected session is
  re-established on the next call.
• A failed write is not automatically replayed: it may already have reached the
  server. Albert does not silently switch to an empty local memory when MCP fails.

---

Upgrading from v0.2.0

Embedded memory remains the default. To use a kaeru MCP server:

[memory]
backend = "mcp"
transport = "http"
url = "http://127.0.0.1:9876/mcp"
token_env = "ALBERT_MEMORY_TOKEN" # omit for an endpoint without authentication
timeout_secs = 30

Set ALBERT_MEMORY_TOKEN in the environment or .env to match the server's token,
then restart Albert. Use HTTPS for a remote endpoint. Switching backends does not
copy existing memories; configure cloud endpoints on the MCP server when using
external memory.

Use kaeru-mcp 0.7.4 or later. If you maintain a custom system.md, update the
old tool names: kaeru_remember → kaeru_episode, kaeru_read → kaeru_at,
kaeru_test → kaeru_evidence, and follow their current argument schemas.

Full configuration, including stdio:
memory backend.

Built on Octo v0.2.0, kaeru v0.7.4 and rig 0.35.

Early and in active testing — feedback and issues are very welcome.

---

Binary

albert-v0.2.1-x86_64-linux-gnu — a release build for x86_64 Linux with glibc 2.39
or newer (e.g. Ubuntu 24.04). It links the system libcurl.so.4, OpenSSL 3 and
libstdc++. Long recordings and videos also need ffmpeg. Deployment:
docs/deploy.md.

SHA-256:

330adbd2955a03be5559cc298b00e34497b74f6f0a917161eef6cbe9056ba839  albert-v0.2.1-x86_64-linux-gnu

Verify: sha256sum -c albert-v0.2.1-x86_64-linux-gnu.sha256

Albert v0.2.0

Albert now hears, speaks and draws — natively, on a ChatGPT subscription, through his
own connectors rather than scripts in between. He reminds you on a real calendar
schedule, takes commands, answers in a much richer Telegram, and every part of him can
run on the provider you choose.

---

Hears and speaks

• Voice messages are heard — of any length. Send one and Albert answers as if you had
  typed it; with no transcription connected, he tells you he can't hear right now.
• Recordings and videos of any length. Long files are cut on their pauses,
  transcribed in parallel and come back with [hh:mm:ss] timecodes — a two-hour meeting
  works as well as a voice note.
• He answers with his voice. A native WebRTC call to ChatGPT Voice turns his reply
  into a voice note. The audio is kept exactly as the model spoke it (no re-encoding), and
  a packet lost on the network is smoothed over instead of making the voice skip. The
  voice is set in speak.toml — from the low, calm cove to the brighter ones.

Draws

• Image generation, natively — gpt-image-2 in one call, about 20 seconds per image.
• Edits from your photos, and transparent backgrounds for logos, stickers and
  cutouts.
• The imagegen skill stays as Albert's prompting guide, so short requests still turn
  into well-built image specs.

Reminders on a real schedule

• "Every weekday at 9", "every Monday", "on the 1st of the month" — a repeating
  reminder becomes one calendar event with a recurrence rule, in your local time, so it
  stays at 9:00 across daylight-saving changes and pops up before every occurrence.
• Chat nudges on a schedule — when you want Albert himself to write to you ("ping me
  here every weekday at 10"), his scheduler now takes calendar schedules in your timezone,
  not only fixed intervals.

Commands

• System commands answer at once, without the model, even mid-reply: /help, and for
  the owner /cancel, /restart and access control.
• Skill commands — /brief, /draw a red fox, /remind, /plan, /watch <link>,
  /settings… A skill declares its command, and /command args runs that skill through
  Albert as a normal turn.
• /help lists what you may run, and the commands appear in Telegram's menu — the
  owner's extra ones only in the owner's chat.

Telegram, reworked

• Replies render as rich messages — formatting, lists, and tables.
• Photo albums arrive as one message; Albert sees the message you reply to.
• Videos and video notes are perceived; voice notes go out as play-in-place
  bubbles, photos as photos.

Swappable organs

• The brain, hearing, speaking and drawing each pick their own provider. Run the
  brain on an API key with subscription voice — or everything on the subscription (for
  example gpt-5.6-terra), sharing one token.
• Voice, transcription and drawing are connector manifests with settings (default
  voice, image defaults, how long recordings are cut). Albert can adjust them himself and
  apply the change with /restart <connector> — the manifest is re-read on restart.

Leaner and safer

• The Docker image drops the Codex CLI and the Python WebRTC stack.
• One owner for the subscription token: refreshed in one place, written atomically, and
  kept 0600 — no script reads it any more.
• A broken connector manifest is skipped with a log line instead of taking Albert down.

---

Upgrading from v0.1.0

• Subscription token. Voice and drawing need a ChatGPT-subscription sign-in:
  albert login (or docker compose run --rm albert login). With auth = "api_key",
  point [subscription] auth_json at the token store.
• Organs are manifests now. config/connectors/{transcribe,speak,imagegen}/*.toml
  turn them on; remove a file to turn one off. The hearing / speaking / imagegen
  switches in albert.toml are gone — voice messages are heard through the transcribe
  organ.
• Skill commands. Your own skills can declare one with command: (and
  command_about:, command_owner:) in their frontmatter.
• Skills retired. tts and transcribe are replaced by the native connectors;
  imagegen now draws through its connector.
• Host packages (systemd installs): ffmpeg for long recordings and videos; outbound
  UDP for voice replies.

Built on Octo v0.2.0 and kaeru v0.7.3.

Early and in active testing — feedback and issues are very welcome.

---

Binary

albert-v0.2.0-x86_64-linux-gnu — a release build for x86_64 Linux with glibc 2.39 or newer
(e.g. Ubuntu 24.04); it links the system libcurl.so.4 (present wherever curl is). For
long recordings and videos the host also needs ffmpeg. Deployment: see
docs/deploy.md.

SHA-256:

e2ac8e6533cd29668c5e9540ebb9029e082fb116ae68fa631088d03f803edb6f  albert-v0.2.0-x86_64-linux-gnu

Verify: sha256sum -c albert-v0.2.0-x86_64-linux-gnu.sha256

Albert v0.1.0

Albert — an always-on AI assistant that grows with you, and stays fast.

Lives in your chat (Telegram), remembers across sessions with kaeru's cognitive-graph memory, reminds you, and acts: files, scripts, calendar, web search, and skills.

In this release:

• Telegram media both ways: perceives photos (albums too), videos/video-notes, and voice; sends photos, voice notes, and rich (Markdown) replies; sees the message a reply quotes.
• Skills, run in the forkd sandbox: video, video-link, tts (speak on the ChatGPT subscription), transcribe, imagegen, and more.
• Owner controls: /cancel stops the current turn (and any script it started), /restart restarts Albert.
• Memory on kaeru 0.7.3; provider-overload turns are retried; two brains — an API key or a ChatGPT subscription.

Early and in active testing.