Most AI agents never learn from their own work. The model is trained once and then frozen, and anything it seems to “remember” lives in prompts, retrieval and memory files. This series starts with the problems before the solutions: why an agent that solved a task yesterday is no better today, why you can’t simply keep fine-tuning, and where reinforcement learning stops helping. Our answer is modified SFT for policy-free learning, a new approach to SFT for continuous learning, and every post builds towards it.
What This Series Covers
Each post opens with a question practitioners run into, explains the mechanism behind it, and compares the options, including reinforcement learning (RL) and RL with verifiable rewards (RLVR) where they apply. Throughout, the approach Eigenform promotes is modified SFT for policy-free learning: fine-tuning small LoRA adapters online on an agent’s own successful attempts, with no reward setting the direction and with safeguards against forgetting.
Where to Start
The problem
- Start with why AI agent memory isn’t learning: memory, RL and our modified SFT compared as ways to make an agent’s experience stick.
The first obstacle
- Then why you can’t keep fine-tuning forever: catastrophic forgetting in continual fine-tuning, and the techniques that reduce it.
Choosing the method
- Why we use modified SFT for policy-free learning instead of a reward: SFT vs RL for agents with sparse, pass/fail feedback.
A practical start
- The simplest form of our modified SFT, a small adapter trained every night on the day’s successful runs: continual learning with LoRA, hacky but good.
When nobody can verify the answer
- How to train when nobody knows the correct answer: machine learning without ground truth, why a learner judged by coherence turns away true new ideas, and how scoring against held-back outcomes first changes that.
