Building Memory Into Your AI Assistant: Sidekick v1.0.4

Release: v1.0.4 · Three-Tier Memory System

One of the fundamental limitations of AI assistants is amnesia. Every session starts from scratch. You repeat yourself. The agent forgets what it did yesterday. Context evaporates.

Sidekick v1.0.4 fixes that.

The Problem With Stateless Agents

A stateless AI assistant is useful for one-off tasks. But a personal assistant — one that manages your inbox, monitors your repos, and handles your calendar — needs to remember things:

What it did this morning
What you asked it to prioritise last week
How you prefer things to be done

Without memory, every autonomous action is context-free. The agent can't learn from past interactions, can't anticipate recurring needs, and can't give you a coherent account of what it's been doing on your behalf.

Three Tiers, Three Time Horizons

v1.0.4 introduces a layered memory architecture designed around how context actually works across time:

Working Memory — Redis

The agent's immediate awareness. Recent events, active tasks, and current session state live here. Fast, ephemeral, and always fresh. This is what lets the agent follow a conversation without losing the thread.

Episodic Memory — Postgres

A daily diary of what happened. Each day, significant events and completed actions are summarised and stored. You can ask your agent "what did you do for me on Thursday?" and get a real answer.

Semantic Memory — pgvector

Long-term pattern recognition. Preferences, habits, and recurring workflows are encoded as vector embeddings. When the agent needs context, it retrieves what's relevant — not just what's recent. This is what enables proactive suggestions and increasingly personalised behaviour over time.

What This Means in Practice

From v1.0.4 onwards, your Sidekick agent:

Maintains coherent multi-turn conversations — follow-ups and references to earlier context resolve correctly
Builds a queryable history — ask it what it did, when, and why
Recognises your patterns — over time, it anticipates recurring tasks and adapts to your preferences
Gets better the longer you use it — the semantic store compounds; the agent's usefulness grows with your history

Automatic for All Users

There's nothing to configure. The memory system is active on all accounts running v1.0.4 or later. It starts accumulating from your next interaction.

See the Three-Tier Memory System feature guide for a full technical breakdown of how each tier works.