System Architecture

Engram decouples intelligence from the interface. It follows a "Headless Host" pattern where the Python kernel runs as a background service, accessible via REST API or CLI.

Presentation Layer

The Interface (Streamlit)

A Streamlit dashboard that visualizes memory and exposes agent controls. It communicates with the kernel purely over HTTP.

Intelligence Layer

The Brain (FastAPI + Ollama)

The Dockerized core at core/brain.py. Calls Ollama directly over HTTP — no LangChain. Autonomous agents run via APScheduler, in-process on FastAPI startup. No separate worker or Redis required.

Llama 3.1

Chat + Agents

Whisper

Voice (Jarvis)

CalendarAgent

15 min

EmailAgent

60 min

Terminal Genie

Shell fix

Spectre

Code review

Git Automator

Commits + PRs

DocSpider

BFS crawler

Memory Layer

Local Persistence — Dual Vector Store

No cloud databases. Personal memory is AES-128 encrypted before it touches Qdrant. Documentation knowledge lives in a separate Qdrant collection, unencrypted for fast RAG retrieval.

Qdrant — second_brain (768-dim, AES-128)

Qdrant — doc_knowledge (768-dim, unencrypted)

SQLite — users.db · matter_registry.db · agent_activity.db

audit_writer — HMAC chain via Unix socket (sole writer)