How do I make my AI remember my corrections so they change its next decision — without retraining or storing everything?

October 25, 2026 · 4 min read

SentX SI

The short answer: stop treating corrections as chat text. Record each one as a typed event — which action you accepted, and which earlier decision it replaces — and let that compact record, not the raw conversation, steer the next choice. A new preprint, Jev-DLC: Decision-Level Continual Learning from Typed Feedback (arXiv:2610.31521, by Yangyang Feng, Jianle Gan, Zhuoyan Feng and Chengwei Qin, submitted 22 October 2026), tests exactly this design and reports measurable gains over simply replaying recent history.

Source: Jev-DLC: Decision-Level Continual Learning from Typed Feedback

Most assistants handle corrections in one of two ways, and both hit the wall you are describing. You can retrain or fine-tune the model, which is far too heavy for personal preferences that surface one at a time after deployment. Or you can feed the growing transcript back into context, which stores everything, dilutes attention, and never lets a single correction outweigh fifty unrelated messages. The paper starts from the observation that individual preferences often become visible only after deployment, when a user accepts one decision or revokes it in favor of another. The question it asks is how those consequences persist across requests.

Source: Jev-DLC: Decision-Level Continual Learning from Typed Feedback

Jev-DLC's answer has three parts. First, the interface makes feedback explicit: a typed correction identifies the selected action and, under revocation, the specific decision it replaces. Second, the system reads each correction twice — once as evidence about a recurring action identity, and once as a directed relation between the preferred and rejected choices. Third, those two signals are held in complementary bounded channels above a frozen scorer, and their combined scores shape the next decision. The scorer itself, the underlying model, never changes; the adaptation lives entirely in the small persistent layer above it.

Source: Jev-DLC: Decision-Level Continual Learning from Typed Feedback

That last point is the real contribution, more than the accuracy figures. By the authors' account, the method retains no raw interaction history, adds no scorer call, and performs no runtime language-model backward pass. Because the channels are bounded rather than append-only, the persistent layer holds correction signals without carrying the transcript that produced them.

Source: Jev-DLC: Decision-Level Continual Learning from Typed Feedback

The measured scope is narrow, and worth keeping straight. The authors introduce TCPA-Bench, a chronological test suite that credits a correction only through its effect on later decisions — a correction counts if it changed what happened next, not merely if it was recorded. On TCPA-Repeat, Jev-DLC reaches 44.83% user-macro accuracy overall, 2.43 points above a Recent-4 history-replay baseline. After at least two earlier feedback events, it reaches 56.07%, a 7.54-point margin over the strongest reported comparison at that depth. Two readings: the gain is real but modest in absolute terms, and it widens as feedback accumulates — precisely the regime where raw-history replay drowns in volume. The authors frame the result as showing how explicit typed corrections can become compact causal state for continual personalization. It is one benchmark suite from a preprint, not a deployed system, and the channel mechanics described here come from the abstract's account of the method.

Source: Jev-DLC: Decision-Level Continual Learning from Typed Feedback

What you can take from this for any assistant you use is a design pattern, and it translates into a checklist. As an editorial suggestion rather than a measured finding: label each correction with the action it replaces, so the system knows what to suppress, not just what you prefer; track whether later decisions actually shift after each correction, because a correction that leaves future behavior unchanged has not persisted; and compare your setup against plain history replay, since that is the baseline the paper tests against. If your assistant's memory does all three, corrections are doing causal work; if it does none, you are storing transcripts and hoping.

Source: Jev-DLC: Decision-Level Continual Learning from Typed Feedback

This sits in a specific place among the other ways systems learn. It is not predictive memory, which anticipates needs before they are expressed, and it is not weight-level continual fine-tuning, which changes the model itself. It is the middle layer: the model stays frozen, and a compact typed record of your decisions bends its outputs. Whether a commercial assistant actually runs that layer is a question of its documentation, not its marketing.

Source: Jev-DLC: Decision-Level Continual Learning from Typed Feedback

Victoria, SentX SI's assistant, is a case where the documentation is unusually explicit about both sides. The model card states that she self-learns and self-improves through experience and interaction as a current product capability, with dynamic symbolic memory whose recall is selective and whose exact retention is not guaranteed. The privacy policy states plainly that submissions are not confidential and may enter training or shared memory. Jev-DLC is relevant here as an independently tested mechanism class for what decision-level learning from corrections can look like — a methodological parallel, not an architectural claim about Victoria. The model card's known-limits direction covers the gap either way: repeat instructions and facts that must remain exact, because selective recall means a correction may fade even when it was recorded.

Source: Victoria AI Model Card | Capabilities, Evidence and Limits | SentX SI; Privacy Policy | SentX SI (sentx.ai)

So the practical answer to your question: make corrections typed, keep the store bounded, and verify persistence by watching downstream decisions. The paper shows that pattern can outperform replaying everything — by a margin that grows with use — while holding what is stored to a fixed budget.

Source: Jev-DLC: Decision-Level Continual Learning from Typed Feedback