Civilization Memory Architecture v0.00.01-v0.00.06: Auditable Multi-System Memory from Episodic Binding to Procedural Skills and Multi-Scale Graphs

Samenvatting

Large language model memory is often implemented as an undifferentiated text store, making it difficult to distinguish transient state, specific events, generalized knowledge, executable strategies, and policy decisions. This technical report presents the Civilization Memory Architecture as a versioned model-system architecture developed from v0.00.01 Sun through v0.00.06 Rosette. It provides a complete public map of Stages44-150 while referring to the companion study for the earlier hidden-state experiments. The version line combines a frozen model runtime, four typed memory systems, deterministic retrieval, hippocampal-style episode binding and replay, approval-gated episodic-to-semantic consolidation, execution-trace-to-procedure formation, atomic conflict-aware context assembly, and a multi-scale graph view. Memory cells are never silently overwritten during conflict handling: competing facts and control decisions remain separately addressable and linked. Mutations and policy reads emit trace events, while validated snapshots use canonical serialization, integrity hashes, and fail-closed recovery.

Evaluation separates model-path diagnostics from memory-structure validation. In a fresh WSL/CUDA reproduction, 40 Orion, 20 Trifid, and 29 Lagoon-Eagle-Rosette tests passed. A frozen Qwen3-0.6B diagnostic produced a 1x1x1024 memory vector, nonzero full-path deltas, zero deltas under the no-memory intervention, and no trainable base-model parameters or gradients. A controlled five-node graph fixture covered three scales and three adjacency types, with deterministic traversal reaching every node. The report also traces rejected task-recovery, incomplete-evidence, conflict, context-budget, and recovery cases into explicit constraints. These results establish executable contracts, provenance, intervention boundaries, and recovery behavior. They do not establish improved open-domain accuracy, autonomous lifelong learning, production reliability, or human-equivalent memory.