Skip to main content

Alchemize Chat

Where it all began. A fully local AI that learns you, dreams about you, and never leaves your machine.

The Ancestor

Before there was a platform, there was a question: What would a personal AI look like if it truly belonged to you?

Alchemize Chat was the answer — a ground-up AI agent system built from scratch on LangGraph, Temporal, and local LLMs. Not a wrapper around an API. The entire pipeline: intent classification, tool routing, multi-step execution, RAG, vision, memory, knowledge graphs — all running on your machine, orchestrated through a flexible architecture designed to grow.

The Architecture

Alchemize Tech Image

Every major subsystem designed and hand-built from scratch:

  • Custom intent classification routing in under 200ms
  • Agentified tool system — easy to add new capabilities
  • Multi-tool chains triggered from a single message
  • Memory architecture that learns who you are over time

Works across web chat, Signal, Drupal, and terminal. The architecture is the achievement — a flexible, extensible pipeline that makes local AI useful for real work.

A deterministic message pipeline:

  • SetFit — intent classification in under 200ms on GPU
  • LangGraph StateGraph — conditional routing, tool execution, side effects
  • Function calling — single-call parameter extraction per tool
  • RAG gating — LlamaIndex + pgvector with semantic reranking
  • Vision routing — Llava 13B for image analysis

Llama 3 70B quantized on Apple Metal. Mistral 7B for fast fallback. ArangoDB knowledge graph. Multi-tool planning with history persistence. Everything containerized via Docker Compose.

The Memory System

Most AI assistants forget you when the conversation ends. Alchemize Chat was different — seven specialized profilers observed every conversation in the background:

  • Psychology, creativity, health, goals
  • Social dynamics and life narrative
  • No delay, no extra prompts

Every night at 4 AM, the system dreams. A consolidation workflow clusters raw observations, resolves conflicts, and builds canonical entities in a knowledge graph. The AI wakes up knowing you better than it did yesterday.

Seven async profilers firing after each response:

  • Archivist — identity and facts
  • Depth Reader — psychological patterns
  • Strategist — goals and planning
  • Chronicler — life narrative
  • Vitals Tracker — health and routines
  • Creator — creative work
  • Mapper — social connections

ArangoDB stores raw entities, canonical concepts, and typed relations across a 12-type ontology. Sleep workflow via Temporal — clustering, consolidation, relation derivation, and structured snapshot injection into future context.

Memory

Under the Hood

An orchestrated multi-service system. Each layer independently built and replaceable.

Temporal.io workflow orchestration
Temporal.io
Workflow orchestration
LangChain / LangGraph agent framework
LangGraph
Agent state machines
Ollama local LLM serving
Ollama
Local LLM serving
Meta Llama 3 language model
Llama 3 70B
Primary reasoning
Hugging Face SetFit classifier
SetFit
Intent classification
LlamaIndex RAG pipeline
LlamaIndex
RAG pipeline
PostgreSQL with pgvector extension
pgvector
Semantic search
ArangoDB multi-model database
ArangoDB
Knowledge graph
LLaVA vision language model
Llava 13B
Vision analysis
Signal messaging integration
Signal
Messaging integration
Apple Metal GPU acceleration
Apple Metal
GPU inference
Docker containerized services
Docker Compose
Containerized services

What Carried Forward

The DNA of Alchemize Chat runs through every layer of the current platform.

Memory
Structured Memory
Seven profilers proved that AI memory needs taxonomy, not just storage. This became how agents accumulate and organize knowledge.
WorkflowIcon
Multi-Agent Orchestration
Temporal workflows coordinating specialized agents. The pattern scaled from one chat system into a full platform orchestration layer.
Sandbox
Local-First Philosophy
Your data belongs on your machine. The platform inherited this as core architecture — local-first, cloud-enabled, never cloud-dependent.

Hibernating, Not Gone

Alchemize Chat isn't retired. It's hibernating. The codebase is intact, the architecture still runs, and the ideas inside it continue to sprout. This isn't an archive — it's a root system. See what it became.

Explore the Platform