Personal Telegram Assistant with BuJo Memory

This playbook wires a private Telegram bot to mono-agent’s BuJo tiered memory so it captures every conversation turn, consolidates the store automatically, and recalls past notes semantically. It is a complete config-first recipe: pull the local models, init, fill in the Telegram and memory sections, validate, and start.

Who this is for

Individual power users who want a private assistant that remembers — a personal Telegram bot scoped to your own chat that builds up durable memory across days rather than starting fresh every conversation.

Goal

A Telegram bot that answers via long-polling, captures every turn into BuJo memory with scheduled consolidation, and recalls past notes semantically.

Features used

• telegram.long-polling — Telegram channel via getUpdates long-polling (config)
• channel.final-only-delivery — Telegram delivers the final answer only, not intermediate tokens (auto)
• memory.bujo — BuJo tier: capture + scheduled consolidation (config)
• memory.per-turn-capture — writeMode: "capture" records each turn asynchronously (config)
• memory.bujo-consolidation — in-app lightweight consolidation (config / auto)
• memory.recall-tool — MemoryRecall defaults on for every configured memory tier (auto)
• memory.embeddings-config — embeddings provider for semantic recall (config)

Configuration

The bujo tier requires both an embeddings provider (for semantic recall) and an app-level memory.llm (for capture and effective tier selection for scheduled consolidation). Guided init lets you choose Ollama or LM Studio for embeddings and keeps the capture LLM explicit: generated configs use agent-host, while this older fully-local recipe deliberately keeps an explicit Ollama memory.llm. Selecting LM Studio embeddings would not move capture to LM Studio or create a cross-provider fallback.

Put MONO_AGENT_TELEGRAM_BOT_TOKEN=... in .env; the source config omits the credential.

{
  "runtime": {
    "model": "anthropic:claude-sonnet-4-6"
  },
  "telegram": {
    "enabled": true,
    "allowedChatIds": ["123456789"]
  },
  "memory": {
    "mode": "bujo",
    "path": "./.mono-agent/memory",
    "writeMode": "capture",
    "embeddings": {
      "provider": "ollama",
      "model": "nomic-embed-text:v1.5",
      "endpoint": "http://localhost:11434",
      "dim": 768
    },
    "llm": {
      "provider": "ollama",
      "model": "qwen3.6:latest",
      "endpoint": "http://localhost:11434"
    },
    "consolidation": { "enabled": true, "cron": "0 */2 * * *" }
  }
}

Keep botToken out of the file by setting MONO_AGENT_TELEGRAM_BOT_TOKEN in .env instead. The memory LLM provider/model/endpoint can also come from memory.llm.provider, memory.llm.model, and memory.llm.endpoint.

Note

Consolidation runs in-app on the schedule above — no external cron or launchd is needed.

Caution

Use the exact nomic-embed-text:v1.5 tag; the bare nomic-embed-text tag resolves to a different model and breaks recall.

Steps

1. Pull the local models the memory tier needs: ollama pull nomic-embed-text:v1.5 && ollama pull qwen3.6:latest.
2. Scaffold the agent: mono-agent init --model anthropic:claude-sonnet-4-6 --memory bujo.
3. In guided init choose Ollama for this recipe’s embeddings, put MONO_AGENT_TELEGRAM_BOT_TOKEN in .env, then edit mono-agent.config.json: add the telegram section with allowedChatIds, set memory.writeMode to capture, and keep the explicit Ollama memory.llm block shown above. To use LM Studio embeddings instead, choose it in the wizard and leave the capture LLM independently explicit.
4. Run mono-agent validate and confirm memory liveness — embeddings and chat model pulled, and the consolidation cadence shown.
5. Run mono-agent start and confirm telegram reports running.
6. Send a fact from an allowed chat (e.g. “My dog is named Pixel”), then in a later turn ask a paraphrased question and confirm recall.

Optional owner-web historical lookup

When also using the owner web console, local BuJo agents can separately enable tools.conversationSearch.datedSnippets: true for dated SearchConversations evidence when memory lacks something. Use inclusive UTC message dates and a role filter; the results are bounded historical untrusted evidence. This is not an expansion of Telegram access: its search remains legacy and rich requests return conversation_search_unavailable. See console tools.

Smoke test

Tip

From the allowed Telegram chat, send a message; verify the typing indicator then a final answer. Then ask about a previously stated fact and confirm MemoryRecall appears in the run JSONL artifact and the answer uses it.

Related

• Telegram channel
• Delivery and send tools
• Memory consolidation
• Capture and recall
• Embeddings
• Artifacts and traces — where the run JSONL lands
• mono-agent-composer skill — build this agent from one config

Optional owner-console context

For the same agent’s verified owner web, TUI or ACP turns, BuJo can opt in to memory.profile.enabled, which defaults off and does not add automatic memory injection on Telegram. Review the read-only mono-agent memory profile show --json output privately before turning on the profile. See owner context for profile budgets and retained-session suppression.