I had a pain point: every time I go into meetings or several meetings at once, especially when you're juggling multiple responsibilities you lose context.
That's why I built my own solution. The app is called "mnelia." The name is a generic play on a brain analogy, since it's connected to my second brain the database that holds my secondary contexts, the ones that don't need my own brainpower.
The system uses the AI subscription providers you already have, plus a recorder native to your OS. The concept is the same as "anarlog," an open-source meeting recorder with transcripts. I use a heavy Whisper model as the weights, so I can transcribe the audio and generate a summary from that transcription. Once it's transcribed, it's saved to my second brain which means I can ask my AI what happened in a given meeting. For example, "What happened during my capstone class meeting?" It can also draft or create diagrams, which feeds my visual learning style.
But what are the guardrails? This chat is connected to my second brain only. The sweet spot is that I have a CLI tool for it that uses just my ai-mem-search, so whatever ai-mem-search returns is the sole basis for the summary it creates. That's my open-source project: build your own AI memory / second brain. The philosophy behind my open-source tools mirrors how the human brain actually retains memories like synaptic pruning, where you cut the unused or old connections while you sleep. (And yes, we humans hallucinate too when we don't sleep and let our brains manage those connections haha.)
Here's the research I drew from: medium.com/@andreworobator/your-agent-needs-a-bedtime-cc594be3f993 (a Medium post), and here's mine: github.com/rambaarde/create-ai-memory. I'm planning to publish mnelia as well, but not yet it needs more improvement.