Building a Cognitive AI Agent with Memory, Perception & Tool Use | LLM Gateway V3 Demo

In this video, I demo a fully functional cognitive AI agent built from scratch — no LangChain,
no LlamaIndex. Just raw Python with a custom Memory → Perception → Decision → Action loop
powered by a multi-provider LLM Gateway.

The agent runs four real queries that test every part of the system:

🔹 Query 1 — Wikipedia Fetch + Artifact Attach
Fetches Claude Shannon's Wikipedia page, stores it as a binary artifact,
then extracts and answers from the raw content.

🔹 Query 2 — Multi-Goal Planning (Tokyo Trip)
Decomposes a complex query into parallel goals — fetches Tokyo activities
AND live weather in separate steps, then synthesizes a final answer.

🔹 Query 3 — Durable Memory Across Runs
Stores "mom's birthday" as a persistent fact in memory.json.
Ask again in a future run — the agent remembers without re-fetching anything.

🔹 Query 4 — Multi-Source Research Synthesis
Searches multiple sources for asyncio best practices, aggregates results,
and writes a structured research file to the sandbox.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🧠 ARCHITECTURE HIGHLIGHTS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
• Memory — keyword-scored persistent facts, preferences & tool outcomes
• Perception — Gemini-pinned goal orchestrator with positional goal identity
• Decision — auto-routed single LLM call (TINY / LARGE tier via gateway)
• Action — pure MCP dispatch with 4KB artifact threshold
• Gateway — routes across 7 providers: Gemini, Groq, Cerebras, NVIDIA,
OpenRouter, GitHub Models, Ollama

No agent framework. No magic. Every component is explicit Python you can read and modify.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📁 Full source code in the description
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

This is part of the EAG V3 (Extensive AI Agents) course — Week 6 Assignment.

#AI #LLM #PythonAI #AgentAI #MCP #LLMGateway #CognitiveAgent #AIDemo #Gemini #GroqAI

Видео Building a Cognitive AI Agent with Memory, Perception & Tool Use | LLM Gateway V3 Demo канала Saikiran Chepa