Replay Buffer Design: Sliding Window or Full Archive?
How large should the replay buffer be for an LLM agent that fine-tunes on its own trajectories? Why a bounded sliding window often beats a full archive.

How large should the replay buffer be for an LLM agent that fine-tunes on its own trajectories? Why a bounded sliding window often beats a full archive.

LoRA stacking merges adapter after adapter into the base and compounds rounding and rank error each generation. Fresh adapters from a frozen base do not.

Data selection for fine-tuning in a self-training loop: training on the best data collapsed in our 13-generation study, while recent data kept improving.

Afterlife helps a continuous learning AI agent develop a unique personality and skills at weight level with LoRA adapters, not just a memory.md file.

Afterlife is an open-source Agent Skill and CLI tool that lets local LLMs nominate memories, train and test LoRA adapters, and activate or roll back updates.

In our 13-generation continual learning study, top-3 historical generation curation collapsed while a K=3 sliding window SFT buffer adapted on 4x less data.

Why can pass@1 fall to zero while an LLM agent gets better at its task? Our 13-generation SFT study, the probe-then-solve mechanism, and how to detect it.

In a 13-generation agent study, a K=3 sliding window SFT loop retains 52.6% of full-history offline composite score and 58.4% of MB freed on 25% of the rows.

Why isn't DeepSeek a continuous learning LLM? Its papers suggest limits of trust, not compute. A narrow agent that measures outcomes can learn today.

Continual learning LoRA for researchers: train a small adapter nightly on the day's successful runs with modified SFT, so the agent improves at its work.