Personal Episodic Identity Substrate: A 9-Dimensional Architecture for Human Memory in the Agent Economy

Personal memory systems for AI agents have converged on a single paradigm: retrieve relevant past events and inject them as context before inference. We argue this conflates two structurally distinct problems: episodic retrieval (what happened) and identity modeling (who you are). This distinction, originating in cognitive psychology [Tulving, 1972, 1983], has been largely absent from the systems literature. We present the Personal Episodic Identity Substrate (PEIS), a 9-dimensional archi- tecture that models not merely what a person experienced but the evolving structure of their identity: active roles, intentional horizon, values, causal decision chains, and afective state. PEIS is implemented as Self, a running system on consumer hardware (Intel MacBook Pro, CPU-only, no GPU) using a local 3.8B-parameter LLM with no model Fine-tuning and no cloud dependency. After 9 days of continuous operation with a single user and 11 simultaneous passive ingestion sources, Self has accumulated 128 long-term memories, 276 entities, 1,612 semantic relations, and 1,213 causal edges, yielding a graph density of 5.84 edges per node with a projected density of 16.32 at 1,000 memories. Direct recall@5 reaches 38%; role-routed recall@5 reaches 36% via implicit domain routing; indirect associative recall@5 remains at 12% as known limitation and active research priority. Self introduces six architectural contributions: (1) behavioral role inference without ex- plicit user declaration; (2) life-event grouping as a narrative episodic unit built from atomic memories without LLM inference; (3) the Minimum Viable Context projection principle for dimension-selective agent grounding; (4) a three-layer staged-commit protocol preventing belief corruption in long-term identity graphs; (5) a privacy layer under patent (P202631047, OEPM, July 2026) operating before LLM inference; and (6) empirical operation across 11 simultaneous passive ingestion sources with zero required user interaction after setup. PEIS is the first system to treat personal identity as a first-class architectural primitive rather than a retrieval by-product.