Spellar 3.0 - AI Meeting companion with cross-meeting memory

Thanks @zrimko — that "different names for the same project across calls" is exactly the gap that pushed us to build the memory layer instead of just better per-call summaries.

Honest answer: the AI handles a fair amount of this semantically, without manual tagging. If your transcripts have enough context — "Project Phoenix, the Q3 launch we discussed last week" — Ask AI treats them as one thread when you query. It's pattern-matching on meaning, not keywords.

Where it gets fragile is when references are terse — "remind me what we said about the launch" with no other anchors. Two cheap lifts solve it:

• Folders per project/client narrow the search scope so terse references resolve correctly.

• A custom-context Template can hold a mini-glossary: "Track decisions about Project Phoenix (a.k.a. Q3 launch, Acme deal, new pricing initiative)." The AI reads that on every summary and normalizes naming for you.

Quick one back: how many parallel projects per client are you usually juggling? That changes whether folders alone are enough or whether the template glossary is worth the 30-second setup.

@shriya_tijare — that "every call treated separately" problem is exactly what we kept hitting as users of other note-takers ourselves. Per-call summaries are fine for the hour after the meeting, useless six weeks later when context is the actual thing you need.

The piece that really amplifies cross-meeting memory is Templates with custom context — you define a per-topic prompt (e.g. "this is a recurring strategy review — track decisions, owners, and any change from prior weeks") and every summary lands in the same structured shape. When you query across meetings later, the AI has consistent, comparable data to work with — not free-form prose — which is what makes recall actually sharp.

Curious: which tools did you try before that left this gap? 🤔 Always useful to know where the bar is set — and whether we're meaningfully past it for your workflow

Thanks @its_maddy_a — context portability across tools is the single biggest blocker for switching, you're definitely not alone.

Where we can help directly: Spellar imports audio files (.m4a, .mp3, Zoom recordings, etc.) and runs the full pipeline on them — transcript, summary, action items — so they land in your account as full meetings, indexed by Ask AI exactly like your live recordings. If you've got the raw audio from any of those tools (most let you download it), that's the fastest path to actually rebuilding memory from your back catalog.

Where we don't help yet: importing summaries (markdown / Notion exports) as first-class meetings. We've thought about it, but text-only imports make Ask AI's recall worse — no transcript anchor for citations. If you've got summaries but not audio, the workaround today is creating a "context" recording — open Spellar, hit record, narrate the key points of past meetings for 2-3 minutes per client/project, let the summary land. Crude but effective for seeding the memory.

Which format is your existing context mostly in — raw audio, transcripts, or just summaries? That'll tell me whether you're 80% of the way there or need the manual workaround.

Thanks @hunter_powabase  — short architecture answer:

Audio: stays on your device by default. Recording, storage, and the first-pass transcript all run locally on Mac/iPhone/iPad. We never see raw audio unless you explicitly opt into cloud transcription (off

by default).

Transcript + summary: stored on our backend for cross-device sync. You pick which AI provider sees the transcript for summarization — Claude, GPT, Gemini, or Perplexity, routed through our gateway. None of

the model providers retain it (zero-retention agreements with each).

Local-only mode: if you turn off cloud AI entirely, on-device transcription + the local app continue to work — you just lose AI-generated summaries and Ask AI. Practical for sensitive calls or air-gapped

workflows.

Quick one back: evaluating for personal use, or for something you'd deploy at GPT-trainer? Compliance constraints (SOC2, BAA, etc.) change the conversation if it's the latter — happy to go deeper.