The framing that got Rowboat to the HN front page — “open-source, local-first alternative to Claude Desktop” — undersells the actual design decision. The interesting bet isn’t that it runs locally or swaps in Ollama. It’s that memory is a persistent, inspectable knowledge graph — email, meetings, Slack, and conversations indexed into an Obsidian-style backlinked store, on disk as plain markdown — instead of context reconstructed on demand by searching transcripts.

That distinction matters more than it looks. Most production agent stacks I’ve worked with treat retrieval as the memory layer: every turn re-queries a vector index and hopes top-k is enough. It holds up until the relationships between facts carry the signal — who owns which thread, what supersedes what, which decision came after which meeting. A backlinked graph makes those edges first-class and durable. You get the KG-enhanced retrieval story without a separate ontology project, because the graph is the memory.

The operational tradeoffs are real, though. Markdown-on-disk is auditable and diffable — a genuine win for governance — but a graph that ingests every email and meeting needs a supersession-and-decay policy or it rots into contradictory edges. Background agents with browser and tool access raise the same blast-radius questions any agentic system does: what’s the confirmation boundary for irreversible actions? The Show HN thread is light on how conflicting facts get resolved, which is exactly where these systems earn or lose trust.

Still, “your memory is a file you can read, edit, and delete” is the right default, and it’s the opposite of where most hosted assistants are heading. If agent memory becomes an inspectable artifact rather than an opaque embedding store, does the vector-DB-as-memory pattern start looking like a temporary hack?