Make Assist remember

[nielsrowinbik]

Problem statement

Every time someone opens Assist, they start from zero. The assistant has no memory of previous conversations, no awareness of what was discussed yesterday, and no sense of continuity. For quick commands this matters less, but for ongoing management tasks, debugging across multiple sessions, or simply picking up where you left off, it is a real limitation.

Community signals

• An August 2024 forum thread "Storing voice assistant (OpenAI) history" explicitly asks how to persist conversation history across Assist sessions, describing the lack of continuity as a core limitation for more complex use cases.
• A March 2025 forum thread "Why does the conversation ID change every 5 minutes?" reveals that Assist's conversation context expires after five minutes, which a user discovered while trying to build a multi-turn interaction. The poster described this as frustrating and limiting.
• The March 2026 post "Too easy to swipe away Assist window while scrolling" lists persistent conversation history as one of its proposed solutions, showing that users naturally reach for session continuity when they encounter the limits of the current dialog.
• In the "AI assistance is a home assistant coding game changer" forum post (June 2025), the author specifically highlights that they prefer Claude over ChatGPT because it "doesn't forget what you said 5 questions ago."
• This feature request is specifically asking for the feature.

Scope & Boundaries

In scope

• Session persistence tied to the authenticated user, so that conversation history is preserved across page loads as well as devices
• Multiple (named) conversations (like ChatGPT or Claude's conversation list) — considered but deferred to a follow-on
• Surfacing the session in the Assist UI so the maintainer can see and scroll previous exchanges

Not in scope

• Shared or multi-user sessions

Foreseen solution

Conversation history is stored server-side and tied to the authenticated HA user. When the user opens Assist, they start a clean session, or can opt to pick up a previous one. This works across devices as long as they are signed in.

Note: this solution expects #158 to be done as it builds on top of the newly introduced UI.

Risks & open questions

• Storage and data model: Conversation history can grow large. Retention limits, pruning strategies, and storage implications need to be defined. Compacting sessions at their conclusion may be a mitigation to explore (thanks @BryanConradHart).
• Privacy: Stored conversations may contain sensitive home configuration details. Users need to understand what is stored and have clear controls over it.
• Context sharing: It is common in apps like ChatGPT and Claude for new sessions to have access to previous sessions. It should be evaluated whether that behaviour should be mirrored, or whether each session should be isolated (or whether this should be contextual). Another option (thanks @BryanConradHart) is to make previous sessions available for RAG rather than loading them into context fully.

Appetite

Medium — 1 cycle. The main investment is the server-side session storage and the user-facing controls. The UI changes are relatively contained if the surface redesign is already done.

Execution issues

No response

Decision log

Date	Decision	Outcome

[BryanConradHart]

As the conversation history grows, each turn uses more input tokens, so that even simple requests become unexpectedly slow and expensive. Typical users aren't going to understand the ramifications of never clearing their session.

What is the strategy for dealing with this?
Some ideas:

• Only the last chat session is loaded into context (not including what the last chat session itself loaded in).
• compact each session at it's conclusion, and only load the compacted sessions into the context.
• make the previous chat sessions available for RAG rather than loading the full history into context.

Idk if those are good ideas or not, but I think context management needs to be considered for this.

[marcinbauer-ohf]

Thanks @BryanConradHart these are great tips for optimization.

[balloob]

Just a note that today we do have history, but that history is automatically cleared when a user has not interacted with that conversation for 5 minutes. This is what powers the ability to ask follow-up questions in voice or having a chat dialog open with multi-turn conversation.

[marcinbauer-ohf]

The main issue is that if I have a conversation today in the Assist dialog, and I just close and re-open it briefly, all trace (as far as the user sees) is gone and there is no way to get it back. An improvement would be to have at least access to those we do store today, and to have them persistent unless explicitly deleted by the user.

[jan-tdy]

What is the strategy for dealing with this? Some ideas:

* Only the last chat session is loaded into context (not including what the last chat session itself loaded in).

* compact each session at it's conclusion, and only load the compacted sessions into the context.

* make the previous chat sessions available for RAG rather than loading the full history into context.

Claude and other AI's have the following system:

• Every night, a summary of what the user is working on, etc is generated
• That means the prompt does not include full history, just "memory".
• But e.g. Claude can access other chats (>make the previous chat sessions available for RAG rather than loading the full history into context.) But only when he needs to read them - e.g., when prompted.

[jan-tdy]

Can we get this at least in some alpha?