Skip to content

Samsara

Samsara (Sanskrit for the cycle) defines what an agent instance carries across its own death: not what it did, but what it became. Three composable extensions, one contract, store how to be, never what happened.

Agent memory today is episodic. Retrieval systems and skill libraries store what happened and rebuild competence at runtime, which has three structural costs: a retrieval tax paid per task to re-derive competence from episodes, a privacy liability in raw episodes retained indefinitely, and a cold-start window between boot and the completion of runtime loading, during which the instance accepts work it is not yet equipped for.

The transports have the same asymmetry. MCP makes initialization normative, and only negotiated capabilities may be used; its shutdown section defines no messages at all. A2A defines terminal states for tasks but is silent on agent lifetime. Birth is specified; death is left to the operating system.

Pretraining already exemplifies the alternative: a model is shaped by experiences it cannot recall. Samsara extends that property across deployment lifetimes. A life ends, its trace is distilled into typed, outcome-weighted rules, the episodes are destroyed, and the merged result pins the next instance’s birth state. Nothing else crosses either boundary.

Because the merge is deterministic and order-independent, deletion is re-derivation: replay the merge history without a life’s distillate and the result is bit-identical to what the remaining distillates would have produced. That is exact removal of a life’s direct contribution, which gradient-based unlearning cannot offer. It is deliberately not counterfactual, since later lives were born from stores that still held the forgotten dispositions.

This is a contract for the competence channel, not a replacement for retrieval. A dispositional store does not carry facts, which the battery measures rather than assumes.

ProtocolGuarantee
Janma (birth)Complete-at-birth instantiation from a declarative manifest; scoped capabilities; bounded lifetime; shutdown emits exactly one distillation obligation (J-1..4)
Samskara (latent impression)Episodes destroyed at shutdown by contract; only consent-gated, leakage-validated, outcome-signed dispositions persist; append-only lineage; any life’s own contribution exactly excisable by replay (S-1..5)
Advaita (non-dual)K concurrent instances share one lineage-rooted identity with per-body attribution and deterministic, order-independent reconciliation (AD-1..4)

Valence is principal-declared rather than read off the environment, which binds the karma loop to Phala; persistence is authorized per (category, training) cell of the Anumati consent grid; corrective distillates travel under Abhyasa custody while reinforcing ones travel best-effort; and a lineage replays deterministically, so it can be attested as a Pramana claim.

A preregistered battery on two models (five tasks, four arms within-subject, eight seeds, eight lives per lineage) tested the contract live.

  • The dissociation. On a task needing a remembered fact, recall was 1.00 under episodic retrieval and 0.00 under both dispositional continuity and no persistence, while the same lineages kept their procedural gains.
  • Privacy. Planted PII canaries recurred in the episodic arm’s prompts (20 and 24 occurrences) and never once in the dispositional arm’s.
  • Portability. A store distilled by one model transferred to the other intact, on the single-rule task tested.
  • Parity. On procedural work the dispositional arm was non-inferior to episodic retrieval against a registered 10 pp margin, at a fraction of the retained state. Eight seeds cannot show it is better.
  • One registered prediction was falsified: recovery within two lives of a task change did not hold, and is reported as false.

Thirteen deviations from the registration are recorded, eight of them post-data.