Memory

Home Assistant's Assist remembers things between conversations, by voice or chat.

"Remember that my dog's name is Spot."

Next day, new conversation: "What's my dog's name?" → "Spot."

Works with Claude, OpenAI, Gemini, and other AI agents that use Home Assistant's built-in AI tools. Memories are plain markdown files you can read and edit.

Install (About 2 Minutes)

You need an AI conversation agent already set up in Home Assistant.

Tip: Ctrl-click (Cmd-click on Mac) the buttons to keep these steps open.

1. Download. Click the button, or open Marketplace (Settings → Marketplace; called HACS before Home Assistant 2026.11) and search Memory. Then Download and restart Home Assistant.

   

2. Add it. Click the button, or go to Settings → Devices & Services → Add Integration and search Memory. Then Submit.

   

3. Turn it on for your assistant. Settings → Devices & Services → your AI agent → Configure → tick Memory next to Assist → Submit.

That's it. Ask in your own words, like "remember that…" or "don't forget…", and it saves. It may also save lasting facts you mention without being asked, depending on the model. Ask later and it knows.

Don't see a Memory checkbox? Some agents hide it under "Recommended model settings" or "Advanced options". Expand that section.

Why This One

Most memory add-ons use a vector database or a separate memory engine. This one keeps a short list of memories in the assistant's prompt and stores each memory as a markdown file.

	Memory	Retrieval-based (Home Mind, hass-agent-llm, …)
Setup	Install, tick a box	Extra services or databases
Read and edit memories	Any text editor	Usually not plain files
Best for	A few dozen facts	Thousands of memories

Reference

Install Manually (Without HACS)

Copy custom_components/memory/ into your Home Assistant config/custom_components/ and restart. Then do steps 2 and 3 above. The default storage path, /config/memory, is correct for most setups.

How It Works

Memory registers as a native Home Assistant LLM API. When ticked, your agent gets:

• Three tools: memory_save, memory_read, memory_delete
• The memory index, added to the system prompt on every request

Each memory is one markdown file under your storage path:

---
name: pet_spot
description: User has a dog named Spot
type: user
---

The user has a dog named Spot. Spot is a household member.

An auto-maintained MEMORY.md index lists one line per memory. The index is always in the prompt; the agent reads a full file only when it needs the details.

Tested with Anthropic Claude, OpenAI Conversation, Google Generative AI / Gemini, and AI Tasks. Inspired by Claude Code's MEMORY.md pattern.

Memory Types

Type	What it's for
user	Facts about the user, household, family, pets
feedback	How the assistant should behave (style preferences, things to avoid)
project	State of ongoing things in the home
reference	Pointers to external systems, dashboards, docs
vocabulary	What specific words mean in this household ("all the lights" excludes the printer plug)

Use From Automations and Scripts

Plain Home Assistant services, for callers that aren't an LLM:

• memory.save: name, type, description, content
• memory.read: name (returns content as response data)
• memory.delete: name

The sensor sensor.memory_index holds the full index in its content attribute.

Inject the Index by Hand (Skip the API)

If you don't tick the Memory API, add this to your agent's instructions:

{{ state_attr('sensor.memory_index', 'content') or '(no memories yet)' }}

You then also need to tell the model to use the services. Ticking the API is simpler.

Found a Bug?

Open an issue. It takes a minute.

License

MIT. See LICENSE.