Chat Analyser

A chat app with persistent memory where you can see and adjust how the context window gets divided between chat history, summaries, and memories from past conversations. Supports Anthropic and OpenAI models. Memory is stored in Qdrant via mem0.

Requirements

• Docker and Docker Compose
• just (optional but handy)
• An OpenAI API key (required regardless of which chat model you use, mem0 uses it for embeddings)
• An Anthropic API key (only if using Claude models)

Installation

Copy the example env file and fill it in:

cp .env.example .env

OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
SECRET_KEY=some-long-random-string

Then start everything:

docker compose up

Or with just:

just dev

just dev starts Docker detached and runs the Tailwind watcher in the foreground. Migrations run automatically on startup.

The app will be at http://localhost:8000.

Usage

Create an account, then configure an agent before starting a conversation. The agent is where you set the model, system prompt, and the context budget fractions.

The fractions divide up the effective context window between:

Fraction	What it controls
chat_history_fraction	Recent message history
summarised_history_fraction	mem0 summaries of older history
relevant_chat_history_fraction	Semantically relevant messages from earlier in the conversation
memory_fraction	Memories retrieved from other conversations
rag_fraction	Retrieved documents

If chat history exceeds its budget the app falls back to summaries and semantic search, which is where tuning these actually matters.

Each response has a details panel showing which memories were fetched, any tool calls, the model's thinking (Claude only), and the exact system prompt and message list that went into the request.

Common tasks

just test
just makemigrations
just migrate
just manage <command>

Without just:

docker compose run --rm web uv run manage.py <command>