Sonnet-class consistency. Haiku-class price. Your agent learns from its environment at runtime — no retraining, no dataset to maintain, no model files to ship.
Hour two of any serious agent session, even the best models start forgetting what they were told, repeating bugs they just fixed, abandoning the user's plan.
Bigger context windows don't fix this. Bigger models don't fix this — they just make the bill worse.
ActiveMem injects small steering vectors into a cheap open base model's forward pass — but only when a learned detector spots the early signature of a known failure mode.
The agent learns from every mistake in its environment. No retraining. No dataset to maintain. No model files to ship.
The technique is grounded in recent interpretability research — vectors in the residual stream that shift a model's behavior without touching its weights. The product is the pipeline that turns observed failures into persistent, retrievable, classifier-gated activation memories, and the managed feed that ships them to your deployment.
Refs. Stolfo et al., ICLR 2025 · Panickssery et al., CAA · Zou et al., RepE · Li et al., ITI