Voice Assistant Long-term Memory
Imagine asking Home Assistant:
“Where’s my car parked?”
“What’s my Wi-Fi password?”
“Remind me where I put the spare keys.”
With Memory Tool, your Voice Assist can now remember and recall information long-term - just like a personal assistant that never forgets.
Two editions are available depending on your needs:
Local Only - Private, offline, lightning fast.
LLM Integrated - Natural, conversational, powered by GPT/Gemini.
My GitHub repository for all dependency files
Edition 1: Local Only
This version runs entirely inside Home Assistant, no LLM or internet connection required.
Features (Local Only)
-
Works offline - fast, private, and secure.
-
SQLite with FTS5 full-text search for quick lookups.
-
Supports set, get, search, and forget, plus TTL, tags, and scopes.
-
Handles duplicates by updating existing memories or creating new ones.
Installation (Local Only)
-
Install and configure the Pyscript integration through HACS.
-
Copy
scripts/memory.pyandscripts/common_utilities.pyinto theconfig/pyscriptfolder. -
Import the
memory_tool_local.yamlblueprint and create an automation from it. -
Dependency files are included in the repository.
-
Restart Home Assistant - the tool will automatically create its SQLite database on first run.
Edition 2: LLM Integrated
This version connects with an LLM (e.g., GPT, Gemini, …) to manage memory during conversations automatically.
Features (LLM Integrated)
-
Interact naturally and conversationally, in any language you choose.
-
The LLM decides when to run set, get, search, or forget.
-
Can refine queries, update existing keys, or create new entries dynamically.
-
Uses the same SQLite + FTS5 backend, with extra flexibility thanks to LLM reasoning.
Installation (LLM Integrated)
-
Install and configure the Pyscript integration through HACS.
-
Copy
scripts/memory.pyinto theconfig/pyscriptfolder. -
Import the
memory_tool_full_llm.yamlblueprint and create a script (do not change the default script name). -
Requires specific usage rules and system policies - see
home_assistant_voice_instructions.mdfor details. -
Dependency files are included in the repository.
-
Restart Home Assistant - the tool will automatically create its SQLite database on first run.
Example: Local vs LLM
Scenario: Saving and retrieving your parking spot
Local Only Edition (exact commands, keyword-based)
You: “Remember office parking spot is in basement 3 near the elevator entrance”
You: “Find office parking spot”
Assist: “in basement 3 near the elevator entrance”
You: “Find parking spot”
Assist: “I found a few candidates: 1. office_parking_spot: in basement… Say a number to open it, or say cancel?”
You: “One”
Assist: “in basement 3 near the elevator entrance”
LLM Integrated Edition (natural language, context-aware)
You: “Remember office parking spot is in basement 3 near the elevator entrance”
You: “Hey, I left my car somewhere… where did I park again?”
Assist: “You told me your car is parked in basement 3 near the elevator entrance”
You: “Actually I moved it to basement 2, column 6”
Assist: “Got it, I’ve updated your parking spot to basement 2, column 6”
Local = precise, offline, reliable.
LLM = flexible, conversational, smarter.
Real-World Use Cases
-
Store and recall Wi-Fi passwords, parking locations, reminders, phone numbers, or any details you want.
-
Save temporary or long-term notes, with TTL configurable up to 10 years - or forever.
Which Edition Should You Choose?
-
Local Only - Best for privacy, speed, and offline reliability.
-
LLM Integrated - Best for natural conversations and smarter handling of memory.
Both editions will continue to be developed in parallel. I’d love feedback from anyone who tries them - especially ideas for new use cases or improvements!