Give it memory

Memory is software, not magic

The model does not permanently absorb your conversations. We will store selected facts ourselves and put relevant memory back into the prompt when needed.

Start with SQLite

SQLite is built into Python and stores the database in a local file.

CREATE TABLE memories
(
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    created_at TEXT NOT NULL DEFAULT CURRENT_TIMESTAMP,
    category TEXT NOT NULL,
    content TEXT NOT NULL
);

import sqlite3

connection = sqlite3.connect("jarvis.db")
connection.execute("""
CREATE TABLE IF NOT EXISTS memories
(
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    created_at TEXT NOT NULL DEFAULT CURRENT_TIMESTAMP,
    category TEXT NOT NULL,
    content TEXT NOT NULL
)
""")
connection.execute(
    "INSERT INTO memories(category, content) VALUES (?, ?)",
    ("preference", "Keep technical answers concise")
)
connection.commit()

Keep memory intentional

A simple first version can have the user explicitly save memory rather than automatically retaining everything. For example:

/remember preference | Use Celsius for temperatures
/remember project | My home server is named atlas

The downloadable starter implements this pattern and retrieves the most recent memories before each model call.

Later: semantic retrieval

Once basic memory works, you can add embeddings and vector search so the controller retrieves memories and documents by meaning rather than only recency. Do that after the simple system is stable.