Answers are grounded in this paper and cite their sources.
Long-lived AI agent sessions accumulate more than conversational history. They also retain tool calls, tool results, temporary execution scaffolding, and provider-specific artifacts that can substantially increase active context requirements without contributing proportionately to future reasoning. Lumina Reforge introduces a provenance-aware, transactional context reconstruction architecture designed to reduce the active working context of persistent AI agents while preserving their durable conversational record and essential continuation state. Unlike conventional summary-based compaction, Reforge reconstructs a bounded execution workbench from durable conversation history, selected execution evidence, and validated continuity checkpoints. Its architecture incorporates durable-state fingerprinting, provenance separation, exact checkpoint binding, transactional replacement, and fail-closed safeguards against stale or invalid reconstruction attempts. Empirical findings Retrospectively documented live evaluations demonstrated estimated total active-context reductions of 62.2% and 76.6%, with a separate owner-operated shakedown achieving an 82.8% reduction in estimated live-history tokens. The durable conversation remained unchanged during the documented reconstructions. Blind continuity testing demonstrated successful recovery of exact durable engineering identifiers and appropriate rejection of unavailable telemetry. It also identified an important limitation: failure to preserve one exact, task-critical detail that existed only in ephemeral execution state. The manuscript documents this negative result and proposes further investigation into repeated-cycle fidelity. These findings support transactional reconstruction as a practical context-management approach for long-lived agents, while not establishing lossless preservation, superiority over conventional compaction, or measured latency and cost improvements. The paper presents the system architecture, implementation principles, retrospective evaluation, limitations, and directions for future research. Research repository: https://github.com/Bino5150/lumina-reforge-research Lumina implementation: https://github.com/Bino5150/lumina Author: Jason “BINO” Malik — Independent Developer, Lumina Project Publication status: Working research preprint, Version 1.0 (2026). Not peer reviewed. Clean the workbench. Keep the blueprint.
No takes yet. Share an insight, caveat, or question.
Jason Malik (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: