Memory & Context

Autoregressive video diffusion models accumulate large attention KV caches during long-form generation. CORAL-KV exploits dependencies between cache layers to predict later keys and values, then quantizes only the residuals.

REMORY augments a textual summary with a bounded sequence of learned soft memory tokens. The approach helps frozen language models approximate full-history behavior while using only 5.2% of the input positions in one reported evaluation.

EngramEdit turns conditional memory from a capacity-expansion mechanism into an editable knowledge interface. It updates shared n-gram embeddings while keeping the Transformer backbone fixed, aiming to generalize edits across expressions without disturbing unrelated capabilities.

A new study shows that sliding-window and linear-attention hybrids excel under different long-context regimes. The authors argue that their behavior is driven by hybrid positional biases, and introduce Sliding-Window Linear Attention as a compromise between local efficiency and extrapolation.

An experiment with galahad-kv shows that a model’s key-value state can be stored on encrypted local NVMe and loaded later without recomputation. The approach does not create a wider attention window, but it offers a practical way to reuse very long histories more efficiently.

Periscope is a training-free inference method that replaces one costly long read with a set of local and document-spanning probes. It identifies answers and supporting evidence while using far less KV-cache memory.

Long-term memory can make LLM agents more personalized, but it can also push them to over-align with a user's past beliefs. MemAdapter dynamically adjusts how much each retrieved memory should influence reasoning.

RealCompanion introduces a longitudinal benchmark for AI companions, evaluating not only whether systems can retrieve past messages but whether they know when the past matters. Its results suggest that memory is needed less often than standard benchmarks imply—and is still difficult to trigger correctly.

HeteroFold targets a key bottleneck in heterogeneous multi-agent LLMs: the receiver repeatedly prefills context that the sender has already processed. The paper reports substantially faster cross-family cache transfer, including a 10.7× speedup over native prefill for one 32K-context direction.

Remembering interaction history does not guarantee that an agent understands the current world or makes progress toward its goal. PoS maintains explicit beliefs at inference time and detects, diagnoses, and recovers from stagnation, cycles, and drift.

MemFold compresses query-relevant textual memories into a fixed set of continuous vectors and optimizes them through the behavior they enable. Its training combines task-level rewards with confidence-gated on-policy distillation from a frozen textual-memory teacher.

WUSH-KV brings data-aware transforms to KV-cache quantization, constructing separate transforms for keys and values from calibration statistics. The paper reports competitive or better results than the OSCAR transform at 2-bit precision across evaluated models and downstream tasks.

Chunked KV-cache compression can make the same information easier or harder to retrieve depending on its position within a compression cycle. The resulting phase sensitivity may be hidden by average long-context scores.

Long-video QA requires more than retrieving relevant moments: it must establish whether an object seen at different times is the same physical instance. Amazon Science’s GEB links such observations into retrievable entity biographies.

Meta researchers introduce Context Language Models, which represent context as a writable file that the model can update without fixed external rules. The paper reports stronger results and lower compute costs across several long-horizon agent tasks.

ByteDance Seed researchers introduce PISA, a pyramid Top-K strategy that narrows attention candidates from coarse to fine. The approach targets the block-selection bottleneck in long-context attention and is designed for O(N log N) complexity.

Most agent memory systems summarize an experience before they know what a future task will require. JitMem keeps raw trajectories and turns them into task-specific memory at read time, making curation more responsive and easier to train.

LatentPort investigates whether a larger hybrid language model can inherit a smaller model’s live inference state without replaying the historical prefix. In a Qwen3.5 4B-to-9B experiment, transferring recurrent and convolutional state alongside translated KV cache substantially narrowed the gap to native 9B inference.

Jev-Mem separates fast memory-management decisions from structured storage and deeper reasoning. On LoCoMo, the architecture improves judged answer quality while reducing both construction time and query latency.

DeepSeek-V4.1-Flash targets the compute, memory, storage, and bandwidth costs created by long-horizon agent workloads. The paper combines a CED architecture, cross-layer KV reuse, FP4 KV caching, and SWA Bounded Replay to substantially reduce cache footprint.

A Johns Hopkins study combines several continual learning components to reduce forgetting across 100 sequential question-answer tasks, raising final average retention from 1.2% to 34.9%.