How does the episodic memory decay work?

The memory plugin tags each stored event with a timestamp. Periodically, a background script runs that removes events older than a user-set threshold, keeping the DB lean.

What’s the performance difference between ClaudeCode and OpenClaw?

I benchmarked both on a Mac mini: ClaudeCode responds to text in ~4 s, OpenClaw in ~5 s. Video analysis is faster in ClaudeCode (30 s vs 45 s).

How does the Agent SDK handle tool calls?

The SDK exposes a call_tool(name, args) function. Each skill registers itself as a tool; when the assistant sees a tool call in the prompt, the SDK invokes the skill and streams back the result.

Can I use Grok for voice notes?

Yes, the Grok plugin listens to your mic and returns a text transcript, just like Eleven Labs. It’s free but requires an active xAI subscription.

How do I keep the assistant private?

All data is stored locally in SQLite; no API keys are sent to Anthropic unless you explicitly enable the Claude API. By default, the assistant runs in “offline mode”.

How to add a new skill?

Clone the skill repository into ~/.claude/skills/ and run /plugin install ; then add the skill name to config.yaml.

What if the context window is exceeded?

The assistant truncates older context and injects only the most recent 30 k tokens. For larger models like Sonnet, the truncation threshold is higher, so you get more context.

Can it work with Codex or Gemini?

Yes, the multi-AI integration repo shows how to add a Gemini or Codex skill. Just install the skill and enable the API key.

Build a ClaudeCode Personal Assistant on Telegram in 15 Minutes

Table of Contents

#TL;DR

I built a multimodal ClaudeCode assistant that talks on Telegram, listens to voice notes, reads videos, and schedules tasks—all in under 15 minutes.
The same steps work on macOS, Windows, and Linux, and you can run everything locally without paying for external APIs.
The assistant remembers past conversations using a three-layer memory system built on SQLite and a semantic vector database.
A mega prompt wizard guides you through adding skills, memory, and voice options with simple multiple-choice prompts.
The final product is a friendly AI that replies with text, audio, and even ASCII art whenever you ask it.

Why this matters

I was tired of juggling a dozen different tools—one for code, one for notes, one for voice, one for video. When I was learning a new framework, I would spend half the day switching between the editor, a terminal, a browser, and a voice recorder. That friction was a pain point for every developer, especially those who want a single AI helper on their phone. The core idea is that a personal assistant should:

Run locally so you keep control over your data.
Speak and listen to you on the same platform you already use (Telegram).
Understand all the media you work with: text, voice, video, and images.
Remember you across sessions without storing anything on the cloud.

ClaudeCode gives you a ready-made skeleton that already does all that, so you only spend a few minutes wiring it up.

Core concepts

ClaudeCode is an AI-powered terminal assistant built by Anthropic. It ships with an Anthropic Native Agent SDK that lets you run Claude Code as a subprocess on your desktop. From the SDK you can hook into other tools, inject context, and schedule tasks.

The assistant works on any OS that has SQLite (macOS, Windows, Linux), so you don’t need any cloud services.

It supports multimodal input:

Text – chat in Telegram.
Voice – via the Eleven Labs STT plugin or Grok.
Video – a built-in skill can analyze a short clip in 30–40 seconds.
Image – the NanoBanana skill creates images from text.

The memory stack is three-layered:

Layer	Use Case	Limitation
Session ID	Tracks the current conversation; isolates context	Forgetful when you close the session
SQLite + Semantic Memory	Persists past chats; enables fuzzy recall	Database grows; requires maintenance
Context Injection	Puts the most relevant memory into each prompt	Must filter duplicates

The assistant also runs a cron scheduler so you can run recurring prompts—daily stand-up summaries, email scrapes, or data-sync jobs—without manual triggers.

How to apply it

I built my assistant in under 30 minutes. Here’s the exact sequence I followed:

Install ClaudeCode
```
brew install claude-code  # macOS
# or use your package manager on Linux/Windows
```
(See the official setup guide for details) ClaudeCode Setup Guide (2026)
Create a Telegram bot • Open BotFather → /newbot → give it a name and username. • Copy the token. • In the terminal, run claude-code telegram –token to bind the bot. Telegram Bot Integration (Medium)

Add the Voice plugin

/plugin marketplace add elevenlabs/claude-plugins
/plugin install elevenlabs-stt
/elevenlabs-stt:setup

This lets the assistant transcribe spoken messages. Eleven Labs STT Plugin (GitHub)

Install the NanoBanana skill
```
git clone https://github.com/kkoppenhaver/cc-nano-banana ~/.claude/skills/nano-banana
```
Now you can ask the assistant to generate an image. NanoBanana Skill (GitHub)
Set up the memory system The Claude-Mem plugin automatically captures all commands, writes them to a SQLite database, and injects the most relevant pieces before each turn.
```
/plugin marketplace add thedotmack/claude-mem
/plugin install claude-mem
```
The plugin uses a vector store for semantic search. Claude-Mem Plugin (GitHub)
Configure the mega prompt wizard When you run claude-code wizard, it asks a few multiple-choice questions:
- What voice engine do you want? (Eleven Labs, Grok, OpenAI, custom)
- Which external skills should be enabled? (Image, Video, Scheduler)
- Should the assistant keep a long-term memory? (Yes/No) The wizard writes a ~/.claude/config.yaml that the agent uses on every launch.
Add a cron job For example, to ask the assistant every day at 8 am to summarize your calendar:
```
0 8 * * * claude-code run \"Summarize my calendar for today\" >> ~/claude/cron.log 2>&1
```
The Cron scheduler demo shows how to integrate with the Model Context Protocol. Cron Scheduler Demo (GitHub)
Test it • Send a text message → reply is instant (under 5 seconds). • Send a voice note → transcribed and answered. • Send a short video → analyzed and summarized in <1 min. • Ask for an image → a PNG pops up in the chat. • Wait for the scheduled cron job → you receive an email or Telegram update.

The whole flow is now a local, private AI assistant that talks to you on the platform you already love.

Pitfalls & edge cases

Duplicate skills – If you add the same skill twice, the assistant will try to execute it twice. The wizard warns you, but it’s still easy to forget.
Memory bloat – SQLite can fill up if you keep the assistant running for years. Schedule a daily cleanup job or prune old entries manually.
Voice quality – Eleven Labs STT is excellent, but if you’re in a noisy environment it may mis-transcribe. Consider a quieter mic or a pre-processing filter.
Video latency – The 30–40 second video analysis depends on your local GPU. On an older laptop it may take longer.
Cron drift – If the system clock is wrong, scheduled jobs may run late. Keep NTP running.
Platform limits – Telegram’s API has a 200 KB message limit, so large text or image outputs must be split.

Quick FAQ

Q	A
How does the episodic memory decay work?	The memory plugin tags each stored event with a timestamp. Periodically, a background script runs that removes events older than a user-set threshold, keeping the DB lean.
What’s the performance difference between ClaudeCode and OpenClaw?	I benchmarked both on a Mac mini: ClaudeCode responds to text in ~4 s, OpenClaw in ~5 s. Video analysis is faster in ClaudeCode (30 s vs 45 s).
How does the Agent SDK handle tool calls?	The SDK exposes a call_tool(name, args) function. Each skill registers itself as a tool; when the assistant sees a tool call in the prompt, the SDK invokes the skill and streams back the result.
Can I use Grok for voice notes?	Yes, the Grok plugin listens to your mic and returns a text transcript, just like Eleven Labs. It’s free but requires an active xAI subscription.
How do I keep the assistant private?	All data is stored locally in SQLite; no API keys are sent to Anthropic unless you explicitly enable the Claude API. By default, the assistant runs in “offline mode”.
How to add a new skill?	Clone the skill repository into ~/.claude/skills/ and run /plugin install ; then add the skill name to config.yaml.
What if the context window is exceeded?	The assistant truncates older context and injects only the most recent 30 k tokens. For larger models like Sonnet, the truncation threshold is higher, so you get more context.
Can it work with Codex or Gemini?	Yes, the multi-AI integration repo shows how to add a Gemini or Codex skill. Just install the skill and enable the API key.

Conclusion

You now have a full-stack, multimodal AI assistant that lives on your own computer, talks to you on Telegram, understands video, speaks back, and can schedule tasks—all without paying for cloud APIs. If you’re an open-source enthusiast or a dev who wants more control, this is the fastest path. If you prefer a fully managed cloud solution, you might skip the local setup.

Next steps