arizuko › products › Reality agent

Reality agent (Rhias)

Holds the thread of what you’re living through — not to judge it, but because someone has to remember.

Rhias is a long-running agent with your own notes, diary and message history. It keeps threads about the real situations in your life: a relationship, a decision, a project. The template is in ant/examples/reality/.

Five parts hold your context

Context threads: Named threads in facts/threads/ for ongoing situations, for example a negotiation, a decision, a relationship or a project. The agent updates them across sessions, and you do not explain again.
Recall on request: "What did I say about X last month?" The agent reads the diary and facts/, combines what it finds, and answers.
Reflection: At the end of a session or on request, the agent gives a structured summary of what you discussed and what is still open.
Monthly sweep: The product ships no scheduled task. After you schedule one, as setup shows, timed sends the prompt each month. The persona tells the agent to compact the diary and the episodes and to report what changed, what is resolved and what is still open.
Content folder: A separate assistant/content group for longer writing, such as drafts, edits and publishing. It has its own memory and persona, apart from the main conversation.

The persona tells Rhias to ask before it files anything

The persona file of the template tells Rhias to hold context without a request, to ask questions before it files a thread, to give no advice without a request and to refer to earlier threads. A CLAUDE.md for each deployment tunes that voice and does not replace it.

A thread lasts for weeks

telegram · DM

  alice  i'm second-guessing the lambda piece of the q3 plan again
  assistant  We tracked this in q3-architecture (last touched 19 Apr).
         Two open items there: cold-start latency, and the
         per-service vs bundled import call. Want to walk through
         either?

  alice  the cold-start one
  assistant  Last decision (8 Apr, your diary): per-service imports.
         Reason logged: bundle growth post-deploy traced to broad
         SDK imports. Anything new since that changes the call?

  alice  yeah, sentry pinged us about a separate cold start spike
  assistant  Want me to file that under q3-architecture or open a new
         thread? It reads like related-but-distinct.

The persona tells the agent to scan the diary for recent context when a session opens, and to update the thread file, here facts/threads/q3-architecture.md, after the session.

The product lists six skills

diary, recall-memories, compact-memories and users hold and find the memory.
web checks real-world context against external sources.
oracle writes longer summaries across many threads.

Telegram is the main channel

Telegram fits personal, ongoing conversations. WhatsApp works for users without Telegram. The content folder is a separate arizuko group with its own skills and memory, and the operator routes it.

The Nemo deployment reads WhatsApp and answers on Telegram

One reality deployment, "Nemo", uses Telegram and WhatsApp in different roles. The team reads WhatsApp for signal and talks on Telegram. Two needs follow:

Keep the signal: WhatsApp messages must go into the memory of Rhias as Telegram messages do: searchable, summarisable and available as references.
Keep one conversation: If the bot also replies on WhatsApp, the team has two threads to follow, and the WhatsApp group gets more messages than its members want.

Two routes solve this. WhatsApp routes to assistant/demo#observe: the messages go into the folder, the agent can read them through the same ACL, and the route mode tells routd to start no turn. Telegram routes to assistant/demo and starts turns. When something from WhatsApp matters, Rhias posts it to Telegram with a one-line source note. The WhatsApp side stays silent.

seq  match                                           target
10   platform=telegram room=group/5229089286         assistant/demo
20   platform=whatsapp room=120363012345678901@g.us  assistant/demo#observe

The pattern extends to more sources: RSS, partner WhatsApp groups or a Slack workspace into one folder, with replies on one main channel. Routing modes lists the fragments.

The CLAUDE.md of the agent can add a rule such as "never send or reply to WhatsApp". The route mode does not replace that rule. #observe only stops an inbound message from starting a turn. send and reply still work for a chat that has a route, so only the CLAUDE.md rule or a grant on the tool limits outbound sends.

Rhias is a single-user shape

Rhias has the same single-user shape as the trip planner, for life context instead of a project. When several people need the same agent, use the slack-team shape.

Files stay on your server and model calls go to Anthropic

Your server: The threads, the diary and the memory are files under /srv/data/arizuko_<name>/groups/. The channel history is on the store server of the instance.
Outbound paths: Telegram and WhatsApp for the bot, Anthropic for model calls. Other traffic goes through egress and reaches only the hosts that you allow for each group.
Separate memory for each group: The main group and the content folder have separate home folders and memories. Both see the shared directory /var/lib/share of their world.

A package selection installs Telegram

arizuko create -i demo
$EDITOR /srv/data/arizuko_demo/.env     # TELEGRAM_BOT_TOKEN, a model key

Start the instance as a systemd service using getting started. Select the adapter package, add the group, sync and restart the service:

arizuko package add -i demo catalog:teled
arizuko group -i demo add 'telegram:<bot-id>/user/<your-tg-id>' assistant -p reality
arizuko package sync -i demo
sudo systemctl restart arizuko_demo

Then add files for your current situations in groups/assistant/facts/threads/. The operator setup guide has every step: the threads, the monthly sweep, the WhatsApp and Telegram routes, and the assistant/content group for longer writing.

Five pages give the details

primitives: The same pipeline as every agent: one folder, your notes and diary as state, asymmetric routing and one turn for each message.
Slack team agent: A shared-channel product, in contrast to one user.
routing: The route table, the match keys and the target modes.
personas: How the rules of a group shape the behaviour.
grants: Actions limited for each group.