[Project] PERMEAR — Memory and Attention Layer for Home Assistant
A few months ago I posted an early version of PERMEAR.
Since then I ended up rewriting most of it.
The original idea was to give Home Assistant conversation agents a persistent memory. While working on it, I spent some time looking at other projects in this area and realized that many of the memory-related problems were already being explored elsewhere.
What kept bothering me wasn't memory itself.
It was noise.
Home Assistant generates an enormous amount of events. If every state change becomes a memory, memory eventually turns into a database of random facts that nobody cares about.
So I started working on a filter that sits before memory.
The idea came from neurophysiology. In the nervous system, most sensory input never reaches conscious attention. It gets filtered first.
PERMEAR now does something similar.
How it works
Every event receives a score based on:
- Novelty
- Priority
- Anomaly
- Context
Most events are ignored.
Some become memories.
A much smaller number become notifications.
Only after that does memory come into play.
Memory is not a log
One of the biggest changes in the rewrite was moving away from permanent memory.
Memories are reinforced when similar things keep happening and gradually decay when they stop being relevant.
Repeated observations become patterns automatically. No LLM is involved in that process.
The goal is to avoid building an ever-growing archive of household trivia and instead keep only what continues to matter.
Current version
- ARAS salience filter
- SQLite-based memory
- Memory reinforcement and decay
- Weekly pattern consolidation
- Automation suggestions based on observed behavior
- Telegram integration
- Home Assistant native architecture (automations, scripts and AI Task)
Current status
Everything is running in my own Home Assistant installation, but this was such a large rewrite that I'm intentionally spending a few weeks testing it before publishing a HACS release.
At this point I'm mostly interested in feedback from people who have experimented with:
- Memory systems
- Long-running agents
- Notification fatigue
- Attention filtering
- Human-in-the-loop automation
Repository
Feedback is very welcome.