We’ve been building in this space too. The big pain we kept hitting is repeated context, messy memory, and bloated prompts.

RAG alone isn’t enough for agent workflows. You need: - memory that evolves (e.g. “my daughter is 10 → now 12”) - entity linking (“my daughter” = Emma = user_42) - context deduping + summarization under token budgets

We ended up building Flumes, an API that turns unstructured user history into structured facts + human-readable summaries. Think memory + graph + packing, all wrapped in one.

If curious: https://www.flumes.ai Would love to chat more if others are solving this differently.