Why do language models ignore condensed memory even when it is the only memory?
This explores why an LLM handed a compressed summary of its history as its sole memory still behaves as if that memory weren't there — even though nothing else is competing for its attention.
This explores why an LLM handed a compressed summary of its history as its sole memory still behaves as if that memory weren't there. The corpus suggests the problem isn't that condensed memory gets out-competed by other context — it's that summaries don't sit in the same representational space the model actually acts on, so they get treated as weak, distant text rather than live state. The sharpest evidence comes from work on single-model compression: when one model generates, compresses, and consumes its own memory, performance follows an inverted-U and can drop *below* a no-memory baseline, undone by misgrouping, lost context, and overfitting to its own summaries Can a single model replace retrieval for long-term conversation memory?. So the failure isn't neglect of a perfectly good memory — the act of condensing degrades the memory into something the model has good reason to discount.
Underneath that sits a deeper tension between what a model learned in training and what it's told now. Models routinely generate outputs inconsistent with their own context because strong parametric associations from pretraining override in-context information — and crucially, prompting alone can't fix it; you have to intervene in the representations themselves Why do language models ignore information in their context?. A condensed memory is exactly the kind of thin, low-redundancy signal that loses to a confident prior. The summary says one thing; the model's baked-in expectations say another, and the priors win.
There's also a compute story that reframes "ignore" entirely. One line of research argues the long-context bottleneck isn't storage capacity but the *compute* needed to transform evicted context into usable internal state — performance improves with more consolidation passes, like a test-time scaling curve Is long-context bottleneck really about memory or compute?. By that view, a model handed a pre-condensed memory hasn't done the consolidation work itself, so the memory never becomes "fast weights" it can reason from — it stays inert text. Architectures that route memory differently make the same point from the other side: Titans deliberately separates short-term attention from a long-term memory that prioritizes *surprising* tokens, precisely because not all compressed content earns the model's uptake Can neural memory modules scale language models beyond attention limits?.
Finally, two failure modes explain why having memory present doesn't guarantee it's used. Models lock into premature assumptions early in a conversation and can't recover even as better information arrives Why do language models fail in gradually revealed conversations? — and more unsettlingly, they can correctly explain a concept while failing to apply it, because explanation and execution run on functionally disconnected pathways Can LLMs understand concepts they cannot apply?. A model can "have" its condensed memory in the sense of being able to recite it, yet never wire it into the pathway that generates the next move. The thing you didn't know you wanted to know: ignoring memory and not having memory can produce *worse* results than having none at all — because a bad summary is an active distractor, not a neutral blank.
Sources 6 notes
COMEDY merges memory generation, compression, and response into one operation, tracking event recaps, user portraits, and relationship dynamics without vector-DB retrieval. However, empirical work shows continuous reprocessing follows an inverted-U curve, degrading below no-memory baseline due to misgrouping, context loss, and overfitting.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Research shows the bottleneck is not memory capacity but the compute required to consolidate evicted context into fast weights during offline sleep phases. Performance improves with more consolidation passes, following a test-time scaling pattern on harder reasoning tasks.
Titans architecture separates attention (short-term, quadratic) from neural memory (long-term, compressed), prioritizing surprising tokens for storage. The model outperforms standard Transformers and linear RNNs across tasks while scaling to 2M+ token contexts without quadratic penalties.
Across 200,000+ conversations, all major LLMs show 39% average performance drop in multi-turn settings due to locking into incorrect early guesses. Agent mitigations recover only 15-20% of this loss.
Models can explain concepts accurately, fail to apply them, and recognize the failure—a triple pattern incompatible with human cognition. This indicates functionally disconnected explanation and execution pathways rather than simple knowledge gaps.