MNEMA: A Memory-Native Episodic-Semantic Architecture for Persistent LLM Agents

Between invocations, contemporary LLM agents do not persist as processes: they persist as markdown files. Widely deployed frameworks (OpenClaw, LangChain, and the CLAUDE.md stack) construct the Agent abstraction around the LLM invocation; persistent state lives in side-files that the model is re-prompted with each cycle but seldom writes to, so within-cycle edits are bookkeeping, not learning, and between cycles nothing changes at all. We call this the file-based identity pattern: it leaves multi-session competence, knowledge accumulation, and goal maintenance with no architectural home. We invert this construction: the agent is the persistent memory system; the LLM is a service it invokes to think. We present MNEMA (Memory-Native Episodic-semantic Architecture for Agents), a three-tier cognitive architecture grounded in cognitive psychology and computational neuroscience: a generic reasoning substrate (a pretrained LLM, replaceable) and two identity-bearing personal tiers, a time-indexed episodic memory and a typed semantic knowledge graph, connected by two substrate-driven, inter-session dynamics: consolidation (episodic-to-semantic distillation) and self-inspection (intra-graph review plus structurally triggered external acquisition). MNEMA is continuous with Tulving's episodic-semantic distinction, Complementary Learning Systems theory, and classical cognitive architectures (Soar, ACT-R), but locates learning in inter-session dynamics rather than runtime decision cycles.