Researchers from Cedars-Sinai Medical Center, the University of Toronto, and collaborators studied how the human brain maintains task context in working memory as goals change. Analyzing neural recordings from 25 treatment-resistant epilepsy patients with surgically implanted electrodes while participants completed rule-based tasks over several minutes, the team found that medial frontal cortex neurons sustained a stable activity pattern to represent a rule even after cues disappeared, while hippocampal neurons generated a flexible rule representation whose format shifted depending on required behavior.Published in Nature Human Behaviour (2026), the work highlights how the brain encodes the same information through patterns of differing stability and flexibility to support rapid behavioral generalization.
- The medial frontal cortex region, including the pre-supplementary motor area (preSMA) and dorsal anterior cingulate cortex (dACC), was studied; all electrode locations were aligned to a brain atlas and color-coded by region.
- Authors proposed embedding durable, time-stable rule representations like those in preSMA/dACC to address generalization failures in large language models and other machine-learning architectures.
- The medial frontal cortex region, including the pre-supplementary motor area (preSMA) and dorsal anterior cingulate cortex (dACC), was studied; all electrode locations were aligned to a brain atlas and color-coded by region.
- Authors proposed embedding durable, time-stable rule representations like those in preSMA/dACC to address generalization failures in large language models and other machine-learning architectures.
Igor Bonifacic writes that users of Anthropic's Claude chatbot can now exercise more granular control over its "memory" feature, which allows the bot to remember personal details and context across conversations. Users can manage these memories through settings on both web and mobile platforms by editing or deleting specific topics, as well as opting in to saving sensitive information like religion or politics.
- Claude's memory is automatically enabled for all users, including those on free plans.
- "Incognito" mode allows users to have chats that are not saved to memory or used for model training.
- Memory can be siloed within specific projects to prevent overwhelming the context window.
- Users can import memories from other inference providers via a dedicated tool in Claude's settings.
- Claude's memory is automatically enabled for all users, including those on free plans.
- "Incognito" mode allows users to have chats that are not saved to memory or used for model training.
- Memory can be siloed within specific projects to prevent overwhelming the context window.
- Users can import memories from other inference providers via a dedicated tool in Claude's settings.
MindMux presents brain.md, an open-source toolkit that provides a persistent memory layer for coding agents by storing project knowledge as plain Markdown files within a repository. This system ensures that critical decisions and constraints are durable across different LLM sessions and machines via version control. A zero-dependency CLI manages the reading and writing of these files to maintain data integrity through an append-only timeline.
- Uses Markdown instead of databases like SQLite to facilitate easier diffing in git history.
- Features a "correct by construction" design that prevents malformed edits by making the CLI the exclusive writer.
- Supports integration with several agents including Claude Code, Codex, Cursor, and Pi.
- Uses Markdown instead of databases like SQLite to facilitate easier diffing in git history.
- Features a "correct by construction" design that prevents malformed edits by making the CLI the exclusive writer.
- Supports integration with several agents including Claude Code, Codex, Cursor, and Pi.
The z386 project aims to implement the Intel 80386 processor core on an FPGA. Recent optimizations have significantly improved its performance, including the implementation of early start memory access and a reduction in branch jump cycles from 9.25 down to six through microcode refinements. These changes resulted in approximately a 39% increase in DOOM benchmark scores, though the project still falls short of original hardware speed and cannot yet boot Windows.
- Implementation of early start memory access for FPGA cores
- Optimization of store queues and branch jump timing
- Significant performance gains in gaming benchmarks
- Implementation of early start memory access for FPGA cores
- Optimization of store queues and branch jump timing
- Significant performance gains in gaming benchmarks
Meta is addressing high DDR5 memory costs by repurposing legacy DDR4 modules from decommissioned servers. Through a custom-developed Vistara ASIC, the company can attach old DDR4 memory to modern servers running AMD EPYC Turin processors that natively only support DDR5. This CXL 2.0 implementation allows for expanded memory capacity by using DDR4 as a secondary, slower tier for cold data while keeping frequent data in fast DDR5.
- Meta's Vistara ASIC bridges legacy DDR4 with modern DDR5 servers
- CXL 2.0 technology enables tiered memory management via NUMA nodes
- Panmnesia offers scalable CXL controller and switch solutions for data centers
- Strategy aims to mitigate rising DRAM prices and hardware costs
- Meta's Vistara ASIC bridges legacy DDR4 with modern DDR5 servers
- CXL 2.0 technology enables tiered memory management via NUMA nodes
- Panmnesia offers scalable CXL controller and switch solutions for data centers
- Strategy aims to mitigate rising DRAM prices and hardware costs
A new study from NYU Langone Health reveals how the brain's hippocampal CA1 region functions as a memory switchboard to balance learning new information with protecting old memories. By using divergent firing patterns, a core group of neurons can manage both incoming and outgoing signals without crossing lines or overwriting existing data. This mechanism helps maintain stability in long-term memory storage while allowing for continuous plasticity during experience.
- The CA1 region acts as a physical hub connecting the CA3 region to the retrosplenial cortex.
- Divergent firing patterns allow shared neurons to serve separate incoming and outgoing communication channels.
- Hub cells remain active during sleep through sharp-wave ripples to facilitate memory consolidation.
- Research provides potential insights into Alzheimer's disease and methods to prevent catastrophic forgetting in artificial intelligence.
- The CA1 region acts as a physical hub connecting the CA3 region to the retrosplenial cortex.
- Divergent firing patterns allow shared neurons to serve separate incoming and outgoing communication channels.
- Hub cells remain active during sleep through sharp-wave ripples to facilitate memory consolidation.
- Research provides potential insights into Alzheimer's disease and methods to prevent catastrophic forgetting in artificial intelligence.
The author compares the performance of an NVIDIA RTX 5090 against Apple Silicon when running large-scale local Large Language Models. While the 5090 offers superior speed for smaller models that fit within its 32GB VRAM, it struggles with massive models that require significantly more memory. In contrast, Apple's Unified Memory Architecture allows Mac Studio users to access much larger pools of memory, making it a more viable platform for running extremely large LLMs like DeepSeek R1.
This paper explores how reinforcement learning agents can use environmental features, termed artifacts, to function as external memory. By formalizing this intuition within a mathematical framework, the authors prove that certain observations can reduce the information required to represent an agent's history. Through experiments with spatial navigation tasks using both Linear Q-learning and Deep Q-Networks (DQN), the study demonstrates that observing paths or landmarks allows agents to achieve higher performance with lower internal computational capacity. Notably, this effect of externalized memory emerges unintentionally through the agent's sensory stream without explicit design for memory usage.
- Formalization of artifacts as observations that encode information about the past.
- The Artifact Reduction Theorem proving environmental artifacts reduce history representation requirements.
- Empirical evidence showing reduced internal capacity needs when spatial paths are visible.
- Observation that externalized memory can emerge implicitly in standard RL agents.
- Implications for agent design, suggesting performance gains may come from environment-agent coevolution rather than just scaling parameters.
- Formalization of artifacts as observations that encode information about the past.
- The Artifact Reduction Theorem proving environmental artifacts reduce history representation requirements.
- Empirical evidence showing reduced internal capacity needs when spatial paths are visible.
- Observation that externalized memory can emerge implicitly in standard RL agents.
- Implications for agent design, suggesting performance gains may come from environment-agent coevolution rather than just scaling parameters.
Researchers from the Chinese Academy of Sciences have identified a new organizational principle within the default mode network (DMN) that explains how it supports both internal thoughts and external perceptions. The study reveals that the DMN is composed of distinct subregions acting as either senders or receivers of information, allowing the brain to flexibly shift between memory-driven thought and sensory perception.
Key findings include:
* Identification of receiver-like subregions that support information integration during perception through stronger connectivity with heteromodal association networks.
* Identification of sender-like subregions that guide memory-based behavior via coupling with sensorimotor systems.
* Evidence that these subdivisions correspond to specific cognitive modes, such as face recognition versus memory-guided decisions.
Key findings include:
* Identification of receiver-like subregions that support information integration during perception through stronger connectivity with heteromodal association networks.
* Identification of sender-like subregions that guide memory-based behavior via coupling with sensorimotor systems.
* Evidence that these subdivisions correspond to specific cognitive modes, such as face recognition versus memory-guided decisions.
Ramp Labs has introduced Latent Briefing, a new method designed to optimize memory sharing within multi-agent systems. By compressing large model KV caches, this approach enables more efficient task decomposition and execution without sacrificing accuracy. Testing on the LongBench v2 benchmark revealed that the solution can reduce token consumption for worker models by up to 65% while actually improving accuracy by 3 percentage points. The technology has proven effective across various document types when tested with Claude Sonnet 4 and Qwen3-14B models.
Key highlights:
- Reduces token usage by up to 65%.
- Improves model accuracy by 3 percentage points on LongBench v2.
- Optimizes multi-agent architectures through KV cache compression.
- Demonstrates faster processing times and high adaptability.
Key highlights:
- Reduces token usage by up to 65%.
- Improves model accuracy by 3 percentage points on LongBench v2.
- Optimizes multi-agent architectures through KV cache compression.
- Demonstrates faster processing times and high adaptability.
