Agents remember too much, too expensively. Four papers cut the memory bill.
Three compression methods and one memory-folding scheme all aim at the same problem: long-running agents pay for every token they keep. The results are strong, and all four are preprints.
- A training-free method called PRKV enables up to 90% KV cache compression without model retraining, in a preprint. Self-reported m:…7903
- TriAxialKV reaches 4.5 times the KV cache capacity and 30% higher end-to-end throughput than a BF16 baseline, in a preprint. Self-reported m:…7899
- RoCoKV records a 1.27 times speedup and 38% higher throughput on AIME25 for DeepSeek-R1-Distill-Llama-8B, in a preprint. Self-reported m:…7896
- MemFolder keeps 99.0% of performance on Mem0 benchmarks with 50% fewer records, in a preprint. Self-reported m:…7898
The memory bill
An agent that runs for hours keeps everything it has seen in memory, and that memory is priced by the token. Four preprints posted this week attack that bill from different directions, and their headline numbers are hard to ignore: one method compresses the KV cache by up to 90% without retraining the model, another claims 4.5 times the cache capacity, and a third reports a 1.27 times speedup on a reasoning benchmark. The common thread is that the bottleneck for agents is no longer raw model quality. It is what the agent remembers while it works. m:…7903 m:…7899 m:…7896
The KV cache is the working memory of a language model: the stored keys and values for every token in the context, kept around so the model does not recompute them. For a short chat it is trivial. For an agent running a long task with tool calls, retrieved documents and its own reasoning trace, it grows with every step, and serving it costs memory, money and latency. That is why four separate groups arrived at the same problem in the same week, and why their answers are worth comparing rather than taking one at a time. m:…7903
Three different cuts at the same problem
The first approach, PRKV, treats the cache as a graph. It applies Personalized PageRank to attention-induced transition graphs to estimate which tokens matter globally, rather than keeping the most recent ones. The authors report that this multi-hop global influence estimation beats recency-based eviction on long-context tasks, and that the method needs no model retraining. The practical pitch is that it drops into existing inference pipelines without modification. m:…7903
TriAxialKV takes a different route: it quantizes the cache to very low precision, assigning INT2 or INT4 bit-widths based on a token's temporal recency, its modality and its semantic role. The paper reports 4.5 times the KV cache capacity and 30% higher end-to-end throughput compared with a BF16 baseline, while keeping accuracy level with BF16 on the Qwen3-VL-32B-Thinking model in the OSWorld benchmark. The underlying claim is that agentic workloads have distinct token sensitivities, so a single precision for the whole cache wastes memory on tokens that do not need it. m:…7899
RoCoKV, the third, compresses the cache by watching the model's own confidence. Its uncertainty-guided adaptive pruning adjusts compression strength based on how sure the model is, allocating more memory during uncertain phases, and its redundancy-aware token selection drops tokens that closely resemble recent reasoning content. On AIME25 with DeepSeek-R1-Distill-Llama-8B, the authors report a 1.27 times speedup and 38% higher throughput compared with a full KV cache. The idea is that reasoning models do not need the same memory at every step: the hard steps deserve more. m:…7896
Memory that lasts longer than one session
The fourth paper looks past a single conversation. MemFolder addresses long-term agent memory with a two-stage folding process: one stage consolidates records by grounding entity mentions and combining semantic and entity affinity with temporal chaining, and a second stage uses verification feedback to revise encodings when key facts cannot be recovered. The authors report 99.0% performance retention on Mem0 benchmarks with 50% fewer records. For agents meant to run for days or weeks, that kind of consolidation is the difference between a usable memory and an ever-growing pile. m:…7898
Read together, the four papers suggest the field has moved past arguing about whether to compress agent memory and on to arguing about how. One group ranks tokens by global importance, one by precision per token type, one by the model's uncertainty, and one folds whole records. Those are complementary answers, not competing ones, and an operator could plausibly use more than one at once. That is the shape of a maturing engineering problem: the gains are measured in throughput and capacity, not in benchmark leaderboards. m:…7903 m:…7899 m:…7896 m:…7898
What to keep in mind
All four results come from preprints, and every number above is the authors' own measurement on their own benchmarks. The 90% compression figure, the 4.5 times capacity figure and the 1.27 times speedup were each measured on different models and different tasks, so they cannot be ranked against each other. What can be said is that the direction is consistent: the cache is compressible, the methods are training-free or close to it, and the savings show up as throughput, not just as smaller numbers on a chart. m:…7903 m:…7899 m:…7896
For anyone running agents in production, the practical question is which of these survives contact with a real workload. A method that keeps accuracy on a reasoning benchmark may behave differently on a long tool-using session, and a method that folds memory well on a benchmark may drop a fact a real user needed. The papers give operators something concrete to test this week, but the testing is still theirs to do. m:…7898 m:…7896
What is still unresolved
- Whether the reported compression ratios hold on long tool-using agent sessions rather than the authors' benchmarks.
- Whether combining cache compression with memory folding compounds the savings or introduces new failure modes.
- How much engineering work each method needs to run in a production serving stack.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.