Can selective history filtering address topic drift that generation-time topic following cannot prevent?

This question reads as a contrast between two places you could intervene against topic drift: filtering the history that goes *into* the model, versus steering the generation that comes *out* of it. The corpus comes down fairly clearly on the input side. The most direct evidence is the finding that automatically selecting relevant previous turns beats stuffing the full conversation into context — and the reason is exactly topic drift: every topic switch a user makes injects irrelevant material that contaminates retrieval, so jointly learning *what to keep* and *what to fetch* outperforms both full-context baselines and even human annotation Does including all conversation history actually help retrieval?. In other words, by the time generation begins, the damage is already done if the wrong history is in the window. Filtering upstream prevents a problem that downstream topic-following can only paper over.

The striking lateral move is that several notes argue the most reliable form of "filtering" is to carry almost no history at all. Markov-style memoryless reasoning decomposes a problem so each step depends only on the current state, never on prior steps — deliberately throwing away accumulated history as "baggage" while preserving answer equivalence Can reasoning systems forget history without losing coherence?. That's selective filtering taken to its logical limit: the safest subset of history is sometimes the empty set.

Why go that far? Because piling up history actively backfires. A single-model memory system that continuously reprocesses event recaps, user portraits, and relationship dynamics follows an inverted-U curve — past a point it drops *below* a no-memory baseline, undone by misgrouping, context loss, and overfitting Can a single model replace retrieval for long-term conversation memory?. So more remembered context is not more grounding; it can be more drift. This is the mechanism that pure generation-time topic-following can't escape: if the context itself is polluted, no amount of on-topic decoding rescues it.

The corpus also shows the complementary defenses on the generation side, which help explain why filtering is the load-bearing move. Grounded-refusal systems constrain generation to only what the evidence supports, trading coverage for integrity when sources are noisy Can RAG systems refuse to answer without reliable evidence?, and gated write-back systems verify outputs through entailment and novelty checks before letting them re-enter the corpus, so drift can't compound across turns Can RAG systems safely learn from their own generated answers?. Both are essentially filters too — one on what generation is allowed to say, one on what gets remembered next — which suggests the real pattern across the collection: drift is fought by curating what flows in and out, not by hoping the decoder stays on track.

The thing you might not have expected to learn: "memory" and "on-topic" can be opposing forces. The intuitive fix for drift is to give the model more context so it remembers the thread — but the corpus repeatedly finds that selectivity, even aggressive forgetting, is what keeps a conversation coherent.