Published
Inherited, Not Remembered: Lifecycle Consolidation for Successor Language Agents
When a newer AI model replaces an older one, it shouldn't just copy all the old conversations — it should keep the useful skills the old model learned from experience while forgetting the specific events, like remembering a lesson without remembering the exact day you learned it.
Abstract
Long-lived language agents increasingly accumulate experience during deployment, yet current model lifecycles provide no principled account of how that experience should be handled when one model replaces another. Continuous updating risks drift and contamination; external memory preserves traces without necessarily producing capabilities; and behavioral transfer does not determine what should be inherited or forgotten. This paper proposes lifecycle consolidation: an explicit phase in which predecessor experience is audited, abstracted into generalizable faculties (experience-derived capacities), and transferred to a successor while episode-specific traces are discarded or archived. We define the episode/faculty distinction, present a six-stage reference architecture, and derive falsifiable comparisons among successors trained from scratch, continuously fine-tuned models, episode-transfer successors, and lifecycle-consolidated successors. The central prediction is a dissociation: a successor retains useful capacities derived from deployment history without recalling their source episodes. Model replacement thereby becomes a problem of selective inheritance rather than simple substitution.
In simple terms
The main idea: keep the lesson, not the diary
When an older AI is replaced by a newer one, it may have a lot of useful experience. The question is what the successor should inherit.
The paper argues that the new model should inherit skills shaped by experience, not the raw memories of exactly what happened.
Why copying memories is risky
If the successor receives old conversation logs, it may carry private user details, noisy mistakes, and brittle examples. It may remember the episode without truly gaining the ability.
Continuous fine-tuning has a different risk: the model may drift, forget older abilities, or absorb contamination from deployment history.
Lifecycle consolidation
The proposed process audits the predecessor's experience, filters out junk and private details, extracts general patterns, turns those patterns into trainable faculties, and transfers those faculties to the successor.
The successor should become better at the task without being able to recall the original episodes that taught the predecessor.
The proof: useful amnesia
The success signature is a dissociation: the successor performs better because it inherited a capacity, but fails when asked to reproduce the original training episodes.
That would show that model replacement can be selective inheritance, not just copying or forgetting.
Keywords
License
Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International