PlatypusDB: A Bitemporal, Agent-Native Database Built for WunderOS

Pentad Labs has detailed PlatypusDB, a purpose-built database designed to live inside WunderOS as its agentic memory substrate. The architecture pairs a Gleam (OTP) control plane with a Zig data plane, running in-process with the agentic OS rather than as an external service. At its core sits a Merkle WAL that serves as the single source of truth, with graph, vector, versioned-tree, and image views all derived from that append-only log.

The design partially agrees with the 'agents just need a filesystem' camp — locality to the execution substrate matters — but argues a filesystem lacks two things agents actually need: bitemporal clock-keeping and fuzzy recall alongside crisp queries. Datalog handles exact retrieval; VSA-based resonance handles approximate recall, both reading the same underlying log. The system replays deterministically and is fuzz-tested with a TigerBeetle-style VOPR.

The accompanying essay makes a pointed case against 'database assemblages' — stitching together Neo4j, Postgres, pgvector, and an audit table via glue code — arguing each component keeps its own clock and its own notion of a fact, so the seams inevitably diverge under audit scrutiny. PlatypusDB instead treats one append-only log as ground truth, with every other view (graph, tree, vector) treated as a redescription rather than a competing source of authority.

For engineers building agentic systems, the relevance is architectural: a single-write-path, dual-clock database eliminates the glue-code reconciliation problem that general-purpose databases push onto application layers, particularly where provable provenance and replay determinism are hard requirements.