Memory & Context

A study of Qwen3.5 and Falcon-H1 finds a sharp functional split between the KV cache and the fixed-size recurrent state. Attention retrieves what appeared in context, while recurrence largely determines the language and persona used in the response.

LatentStream presents a latent working-memory framework for streaming video understanding. Instead of repeatedly attaching retrieved visual evidence as external context, it progressively internalizes that evidence into a compact, evolving representation.

Random Attention shows that long reasoning traces may not require token-level importance scoring. By protecting the prompt and evicting other entries uniformly at random within each attention head, the method approaches strong learned selectors while improving serving throughput.

LatentPress compresses conversations and long documents into continuous memory tokens that a frozen decoder can consume through its input-embedding interface. The approach avoids text reconstruction and reports competitive accuracy with substantially lower context processing cost.

VoiceMem separates factual recall from emotional and persona modeling in a streaming memory architecture for duplex speech language models.

Anthropic is merging the memory systems behind Claude Chat and Cowork, allowing users to move from discussion to execution without repeating project context. The update also adds controls for reviewing, editing, and deleting saved memories.

Remembering more does not always help a language model solve the task in front of it. MemTrapBench measures two ways retrieved memories can distort reasoning and introduces AdaptiveMem as an inference-time mitigation.

A unified study compares a broad range of memory substrates for memory-augmented LLM agents. Its findings suggest that the best substrate depends on the task, history length, and efficiency regime rather than on a single global ranking.

OpenAI is adding Computer History to the ChatGPT desktop app on macOS, letting ChatGPT and Codex reference a timeline of user actions. The feature avoids screenshots, but it still raises familiar questions about how much of a personal computer an AI assistant should be allowed to remember.

Agent Memory Distillation reframes knowledge transfer for tool-using agents: instead of retraining, it distills successful teacher trajectories into layered memories that small models can reuse. The goal is to make weak agents act more reliably without collecting massive new data.

A new paper identifies a subtle failure mode in ALiBi positional encoding: linear attention biases can underflow in floating-point arithmetic, zeroing out many attention weights. The result is that some heads become partially blind to distant tokens.

A new paper frames cross-session personalized ambiguity adaptation as a benchmark task for coding assistants. The goal is to see whether models can use a user’s resolved history to write the intended code with fewer repeated clarifications.

The paper scales Memory Decoder to 6.9B parameters and 300B training tokens, arguing that a small backbone paired with a large parametric memory can be more parameter-efficient than simply enlarging the base model.

Metis introduces the idea of a memory foundation model, moving agent memory from external retrieval pipelines into a persistent state inside the model backbone. The result is a prototype where history can be compressed, updated, and accessed without replaying the full context.

CodeNib addresses a common bottleneck for coding agents: repeatedly rediscovering the same repository context. It builds lexical, dense, and structural views per commit and serves search, navigation, and bounded context through one runtime.

The paper introduces Agentic Context Management, arguing that production agent failures often stem from poor control over what sits in the reasoning context, not from weak reasoning alone.

ReflectWorld-MM is designed to help AI systems remember open-ended video streams the way humans do: by tracking persistent people, objects, and changes over time. Its main shift is from frame-based storage to entity-oriented memory.

Long context is useful only when the right information survives. SWE-Pruner Pro argues that a coding agent already has enough internal signal to decide what tool output should stay or go.

RecGPT-V3 presents an industrial LLM-based recommender that maintains evolving user memory, grounds intent with Semantic IDs, and compresses reasoning into latent tokens. In Taobao’s “Guess What You Like” feed, the report cites higher engagement and a 52.4% reduction in serving resource consumption.

xHC revisits Hyper-Connections as a practical scaling axis for large language model pre-training. By expanding residual streams to N=16 while sparsely updating only part of them, it reports stronger downstream performance with modest added training cost.

A new paper proposes storing verified knowledge as byte-exact KV-cache artifacts and grafting them back into a frozen small language model. The result is striking efficiency on repeated cases, but the claims also raise questions about reproducibility and what counts as improved reasoning.