Samsara: Born With It. Dispositional Continuity, Turnkey Instantiation, and Substrate Identity for Ephemeral Agent Instances

Agent memory today is episodic: retrieval systems and skill libraries store what happened and rebuild competence at runtime, paying that cost per task and keeping raw episodes indefinitely. We specify the Samsara Layer, three composable protocol extensions that invert it. Samskara compresses a working life into typed, outcome-weighted rules at shutdown and destroys the episodes. Janma starts an instance complete, from a declarative manifest. Advaita lets many copies of oneagent run at once under one identity, with each action traceable to the copy that took it, and merges what they learn to the same result in whatever order it arrives. Because that merge is deterministic, deleting a past life means replaying the merge without it, which removes what that life contributed exactly rather than approximately. Thirteen safety invariants are specified here and enforced at runtime in a Python implementation. A preregistered battery on two models (five tasks, four arms, eight seeds, eight lives) tests this. On a task needing a remembered fact, the episodic arm recalled it every time and the dispositional arm never did, scoring exactly like an agent with no memory at all, while those same agents kept the skill they had built. Planted personal-data strings recurred in the episodic arm’s prompts and never in the dispositional arm’s. A store built by one model worked unchanged in the other. On procedural work the dispositional arm was statistically non-inferior to episodic retrieval while keeping far less data, though only for the tasks tested and the single retrieval comparator run, and eight seeds cannot show it is better. One registered prediction, recovery within two lives of a task change, was false. This is a prototype and a first measurement, not a deployment result.