Positioning Thesis
Agent memory should feel like infrastructure, not a hack
The core Memnode claim is simple: long-term memory for agents is closer to a database engine than to a prompt pattern. If the underlying system cannot track provenance, isolate tenants, and survive restarts, it is not serious memory.
Memory is a database problem
The failure mode in agent memory is usually not "the LLM forgot the prompt". It is missing persistence, provenance, conflict handling, and lifecycle controls.
Provenance beats vague retrieval
Observed, reported, inferred, and hypothesized memories should not be mixed together. Memnode keeps those distinctions visible.
Hosted plus local is the right shape
Purely local tools are hard to share. Purely hosted tools are hard to trust. Memnode keeps the local MCP path while adding a real hosted SaaS surface.
Boring operations matter
Signed leases, quota enforcement, tenant caches, and a Rust data plane are not glamorous, but they are what make hosted memory usable.
The product tradeoff
The SaaS is intentionally simple: Supabase for auth and billing, Cloudflare for dashboard and control plane, and one shared Rust data plane per region. That is enough to turn Memnode from a local tool into a product without pretending the first version is a hyperscale platform.