Ask HN: How are you solving long-term memory for production AI agents in 2026?

Developers building production AI agents in 2026 are using simple vector search, keywords, BM25, text matching, and RRF for long-term memory, avoiding graph construction due to costs. One team reports a single SQLite file works for up to a few million document chunks.

Specifically interested in teams who moved past demos into real production workloads. Mem0, Zep, custom solutions — what's actually working and what keeps breaking? Simple vector search + keywords + bm25 + text match + RRF. We specifically avoided graph construction due to associated costs. Everything is in just one sqlite file. Works fine for up to a few million document chunks.