From Execution Trajectories to Durable Cognition: Closing the Lifelong Self-Improvement Loop in Resource-Constrained Autonomous Agents

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Autonomous Large Language Model (LLM) agents operating in local workstation and edge environments are fundamentally limitedby physical hardware: finite Key-Value (KV) cache memory (≤ 16GB VRAM), narrow context windows (8K–32K tokens), andgeneration compute ceilings. Under these constraints, an agent’s inability to learn from runtime execution produces an acute failuremode: Operational Amnesia. Identical compiler errors, OS shell traps, and tool misconfigurations are repeatedly encountered,investigated, and debugged across independent sessions, dissipating up to 61.4% of prompt tokens and 66.7% of conversational turnson redundant trial-and-error cycles. Existing memory architectures fail to resolve this dilemma due to multi-day batch consolidationlatency (the Amnesia Gate), query-agnostic retrieval that pollutes finite context, and unconditional retrieval reinforcement thatentrenches irrelevant knowledge in stagnation echo chambers.In this paper, we introduce the theoretical foundations and empirical realization of the IronClad Continuous LearningSubsystem, a Rust-native autonomous cognitive architecture engineered for lifelong self-improvement on consumer hardware.IronClad introduces five coordinated systems: (1) Immediate Post-Task Distillation, extracting runtime friction and recoverystrategies within ≤ 2 ms upon task termination; (2) Dynamic Task-Aware Retrieval, projecting incoming goals into a domainkeyword topology to deliver surgical context without token bloat; (3) Outcome-Grounded Reinforcement Scoring, tying memoryimportance updates to verified empirical task utility rather than retrieval frequency; (4) Dual-End Episodic Trajectory Sampling,preserving both initial intent and terminal resolutions within an O(1) token footprint; and (5) Autonomous Direct PreferenceOptimization (DPO) Trajectory Synthesis, streaming verified trajectories into ShareGPT datasets for downstream local parameteradaptation (QLoRA). We analyze the convergence properties of the Monotonic Error Suppression and Bounded ImportanceEquilibrium mechanisms under explicitly stated operational assumptions. Empirical evaluations on 20 sequential realistic compilationand refactoring tasks demonstrate that IronClad collapses recurrent error rates from 64.6% to 4.2%, cuts mean turns to resolutionby 66.7%, reduces prompt token consumption by 61.4%, and achieves a 3.4× end-to-end wall-clock speedup. Furthermore,on a strictly held-out test suite (n = 50), autonomous local QLoRA fine-tuning on IronClad-synthesized trajectories executes in12.0 minutes on consumer hardware (RTX 5060 Ti), elevating valid AST edit rates from 88.0% to 90.0% (+2.0 percentage points,representing 1 additional task out of 50 within statistical noise) while establishing the empirical scaling boundaries of parameteradaptation versus in-context procedural distillation