Unabyss and ZooData top Product Hunt back to back

Product Hunt crowned agent memory infrastructure its top product two days running, per the July 17-18 leaderboard.

Unabyss took the #1 spot with 598 votes on July 17, and ZooData followed with 606 votes on July 18. Neither is an app anyone opens; both are the layer other agent products plug into, which is where switching cost accumulates.

One day like that is noise. Two days running, with the winner being infrastructure rather than an application, is worth a second look.

If Claude Code disappeared tomorrow, moving to Codex is a day of work: porting skill files, adjusting prompts. Losing eighteen months of accumulated project context, corrections, and preferences is a different problem. That's not portable at all.

It's Segment's playbook, one layer up. Segment won because once every downstream tool depended on your schema, ripping it out meant re-plumbing everything. Memory has the same shape, except what's stored isn't routed data, it's accumulated judgment about you.

The research side caught up. Same week. A paper on arXiv (2607.05794) describes NapMem, a memory system built in layered tiers: raw conversation, typed records, topic tracks, profiles.

The agent is trained via reinforcement learning to pick which tier to query. It holds up on PersonaMem-v2, LongMemEval, and LoCoMo without wrecking the general reasoning that memory-specialized training usually breaks.

The design lesson survives without the RL training. Expose memory as several typed tools at different granularities, and let the model choose which to query instead of running one search over everything.

I've been running a single vector store that only grows. Retrieval quality drops exactly the way you'd expect, once everything in the corpus looks somewhat similar to everything else.

I don't know yet whether Unabyss or ZooData ends up being the one that matters. I'm confident the category does.