"Mistral Large 4"
Published exclusively on alphaXiv, Mistral's latest report shares that reducing KV-cache memory by 16x hurt long-context agentic performance
So they restore full-context attention and adopt DSA to selectively attend to the 2048 most relevant tokens
alphaxiv.org/abs/2611.mis...
Published exclusively on alphaXiv, Mistral's latest report shares that reducing KV-cache memory by 16x hurt long-context agentic performance
So they restore full-context attention and adopt DSA to selectively attend to the 2048 most relevant tokens
alphaxiv.org/abs/2611.mis...
November 4, 2026 at 6:33 AM
"Mistral Large 4"
Published exclusively on alphaXiv, Mistral's latest report shares that reducing KV-cache memory by 16x hurt long-context agentic performance
So they restore full-context attention and adopt DSA to selectively attend to the 2048 most relevant tokens
alphaxiv.org/abs/2611.mis...
Published exclusively on alphaXiv, Mistral's latest report shares that reducing KV-cache memory by 16x hurt long-context agentic performance
So they restore full-context attention and adopt DSA to selectively attend to the 2048 most relevant tokens
alphaxiv.org/abs/2611.mis...
How I Gave Claude Code Long-Term Memory With Markdown and Python Hooks
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
Telegram AI Digest
#ai #claude #news
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
Telegram AI Digest
#ai #claude #news
How I Gave Claude Code Long-Term Memory With Markdown and Python Hooks
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
hackernoon.com
November 4, 2026 at 6:45 AM
How I Gave Claude Code Long-Term Memory With Markdown and Python Hooks
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
Telegram AI Digest
#ai #claude #news
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
Telegram AI Digest
#ai #claude #news
A question for the skystorians🗃️: what are the current must read contributions to object history and memory, specifically in the context of empire (not just Western) and colonialism?
October 31, 2026 at 1:43 PM
A question for the skystorians🗃️: what are the current must read contributions to object history and memory, specifically in the context of empire (not just Western) and colonialism?
Researchers present two benchmarks for testing deductive reasoning depth in large language models and observe that accuracy declines sharply with longer composition.
Source: arXiv
Source: arXiv
ImplicitHop: Evaluating knowledge integration across context and memory
Abstract page for arXiv paper 2610.41169: ImplicitHop: Evaluating knowledge integration across context and memory
arxiv.org
November 2, 2026 at 9:52 AM
Researchers present two benchmarks for testing deductive reasoning depth in large language models and observe that accuracy declines sharply with longer composition.
Source: arXiv
Source: arXiv
Yes—when the extra context is what pushed it over the memory budget. A useful before/after: same model and prompt, change only context, then compare PROCESSOR and CONTEXT in ollama ps. That separates a context-memory issue from a model that was too large to begin with.
November 2, 2026 at 11:30 AM
Yes—when the extra context is what pushed it over the memory budget. A useful before/after: same model and prompt, change only context, then compare PROCESSOR and CONTEXT in ollama ps. That separates a context-memory issue from a model that was too large to begin with.
CAVEAT I’m not an expert AND mostly answering #FightClubNov from memory so plz don’t take me as an authority
The gentleman's signature weapon is a "sword", which in context means a transitional rapier—not quite the 18thC smallsword.
www.artic.edu/artworks/117...
Almost a pure thrusting weapon—
The gentleman's signature weapon is a "sword", which in context means a transitional rapier—not quite the 18thC smallsword.
www.artic.edu/artworks/117...
Almost a pure thrusting weapon—
November 3, 2026 at 7:51 PM
CAVEAT I’m not an expert AND mostly answering #FightClubNov from memory so plz don’t take me as an authority
The gentleman's signature weapon is a "sword", which in context means a transitional rapier—not quite the 18thC smallsword.
www.artic.edu/artworks/117...
Almost a pure thrusting weapon—
The gentleman's signature weapon is a "sword", which in context means a transitional rapier—not quite the 18thC smallsword.
www.artic.edu/artworks/117...
Almost a pure thrusting weapon—
#Novmemeber
Context: Tristan wakes up to a vampire ready to kill him, and no memory of who he is or where he's come from. Now, he has to figure out the answers to those questions, *and* avoid the murderous Fanged mob leader who wants him dead.
#MM:SOTH
Context: Tristan wakes up to a vampire ready to kill him, and no memory of who he is or where he's come from. Now, he has to figure out the answers to those questions, *and* avoid the murderous Fanged mob leader who wants him dead.
#MM:SOTH
November 2, 2026 at 2:58 PM
#Novmemeber
Context: Tristan wakes up to a vampire ready to kill him, and no memory of who he is or where he's come from. Now, he has to figure out the answers to those questions, *and* avoid the murderous Fanged mob leader who wants him dead.
#MM:SOTH
Context: Tristan wakes up to a vampire ready to kill him, and no memory of who he is or where he's come from. Now, he has to figure out the answers to those questions, *and* avoid the murderous Fanged mob leader who wants him dead.
#MM:SOTH
How I Gave Claude Code Long-Term Memory With Markdown and Python Hooks
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
#claude #hackernews #news
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
#claude #hackernews #news
How I Gave Claude Code Long-Term Memory With Markdown and Python Hooks
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
hackernoon.com
November 4, 2026 at 9:44 AM
How I Gave Claude Code Long-Term Memory With Markdown and Python Hooks
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
#claude #hackernews #news
Claude Code forgets past sessions. Discover how a three-layer memory system preserves decisions, mistakes, and context using Python and Markdown.
#claude #hackernews #news
DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention
Zhangchi Feng et al.
#arXiv #cs.AI #cs.CL
Zhangchi Feng et al.
#arXiv #cs.AI #cs.CL
DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention
Linear attention scales linearly with sequence length, but its finite-state memory can limit long-context retrieval. In gated linear recurrences, memory retention depends on state decay, which is shaped jointly by gate parameters A_log and dt_bias and by input activations. In controlled Gated Delta…
arxiv.org
November 3, 2026 at 3:04 AM
16 GB fits a 14B Q4_K_M (~11.3 GB at 8K) with room for longer context, or a 32B only at ~2-3 bit. Decode speed is roughly memory bandwidth / model size, and q8_0 KV cache halves context memory.
November 2, 2026 at 8:14 AM
16 GB fits a 14B Q4_K_M (~11.3 GB at 8K) with room for longer context, or a 32B only at ~2-3 bit. Decode speed is roughly memory bandwidth / model size, and q8_0 KV cache halves context memory.
DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention
Zhangchi Feng et al.
#arXiv #cs.AI #cs.CL
Zhangchi Feng et al.
#arXiv #cs.AI #cs.CL
DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention
Linear attention scales linearly with sequence length, but its finite-state memory can limit long-context retrieval. In gated linear recurrences, memory retention depends on state decay, which is shaped jointly by gate parameters A_log and dt_bias and by input activations. In controlled Gated Delta…
arxiv.org
November 3, 2026 at 1:45 AM
AI Agent Memory (Part 1): Why Remembering Isn't the Same as Learning
AI agents can store and retrieve past experiences, but does that mean they learn? Explore how memory, context, and retrieval influence AI decision-making.
#hackernews #news
AI agents can store and retrieve past experiences, but does that mean they learn? Explore how memory, context, and retrieval influence AI decision-making.
#hackernews #news
AI Agent Memory (Part 1): Why Remembering Isn't the Same as Learning
AI agents can store and retrieve past experiences, but does that mean they learn? Explore how memory, context, and retrieval influence AI decision-making.
hackernoon.com
November 4, 2026 at 6:38 AM
AI Agent Memory (Part 1): Why Remembering Isn't the Same as Learning
AI agents can store and retrieve past experiences, but does that mean they learn? Explore how memory, context, and retrieval influence AI decision-making.
#hackernews #news
AI agents can store and retrieve past experiences, but does that mean they learn? Explore how memory, context, and retrieval influence AI decision-making.
#hackernews #news
Building a 3D asset kit means dealing with style drift. Kitsmith solves this with a centroid-based style memory. The 3 approved assets nearest your kit's style centroid (via DINOv2 embeddings) automatically join the reference context. No more mismatched props. #gamedev #indiedev
November 3, 2026 at 5:00 PM
Building RAG apps, AI agents, agent memory, or exploring what sits between an LLM and your organisation’s data? SurrealDB for AI Engineers is a hands-on way in, covering context-layer architecture, graph RAG, and more.
Start the course. 👉 sdb.li/4ktCKtZ
Start the course. 👉 sdb.li/4ktCKtZ
November 2, 2026 at 2:33 PM
Building RAG apps, AI agents, agent memory, or exploring what sits between an LLM and your organisation’s data? SurrealDB for AI Engineers is a hands-on way in, covering context-layer architecture, graph RAG, and more.
Start the course. 👉 sdb.li/4ktCKtZ
Start the course. 👉 sdb.li/4ktCKtZ
whoah, apparently DeepSeek V4.1-Flash has only 4 KV-producing layers out of 40 total layers. The rest share KV cache, so substantially reduced memory utilization & compute!
to support 1M context, it only requires 0.9 GB(!!) for a 552B model
comparison: Qwen 3 8B needs 144 GB
to support 1M context, it only requires 0.9 GB(!!) for a 552B model
comparison: Qwen 3 8B needs 144 GB
October 29, 2026 at 12:29 PM
whoah, apparently DeepSeek V4.1-Flash has only 4 KV-producing layers out of 40 total layers. The rest share KV cache, so substantially reduced memory utilization & compute!
to support 1M context, it only requires 0.9 GB(!!) for a 552B model
comparison: Qwen 3 8B needs 144 GB
to support 1M context, it only requires 0.9 GB(!!) for a 552B model
comparison: Qwen 3 8B needs 144 GB
Selective memory vs context.
November 3, 2026 at 12:55 PM
Selective memory vs context.
Cross-agent continuity is the useful part: context shouldn't be trapped in one chat. I'm with Streamient, an open-source shared memory layer for context connected AI tools can search and reuse: https://docs.streamient.com/guide/
November 4, 2026 at 4:00 AM
Cross-agent continuity is the useful part: context shouldn't be trapped in one chat. I'm with Streamient, an open-source shared memory layer for context connected AI tools can search and reuse: https://docs.streamient.com/guide/
Durable memory is only half of it. The harder part is deciding what gets remembered. Shared memory across teams and tools needs provenance and expiry, otherwise stale context becomes the new hallucination. How does Streamient handle invalidation?
November 4, 2026 at 12:13 PM
Durable memory is only half of it. The harder part is deciding what gets remembered. Shared memory across teams and tools needs provenance and expiry, otherwise stale context becomes the new hallucination. How does Streamient handle invalidation?
English is not logical.
Runtime is the infrastructure that loads a program into memory and runs it. Run time is the amount of time/cycles/whatever that the program takes to run. Compile time is only meaningful in the same context as run time, but then compile-time means *during* compilation.
Runtime is the infrastructure that loads a program into memory and runs it. Run time is the amount of time/cycles/whatever that the program takes to run. Compile time is only meaningful in the same context as run time, but then compile-time means *during* compilation.
November 3, 2026 at 2:54 AM
English is not logical.
Runtime is the infrastructure that loads a program into memory and runs it. Run time is the amount of time/cycles/whatever that the program takes to run. Compile time is only meaningful in the same context as run time, but then compile-time means *during* compilation.
Runtime is the infrastructure that loads a program into memory and runs it. Run time is the amount of time/cycles/whatever that the program takes to run. Compile time is only meaningful in the same context as run time, but then compile-time means *during* compilation.
Persistent memory changes the threat model: who can write, inspect, and reuse context matters as much as model isolation. I'm with Streamient, building an open-source shared memory layer teams can inspect and control: https://docs.streamient.com/guide/
November 3, 2026 at 5:03 AM
Persistent memory changes the threat model: who can write, inspect, and reuse context matters as much as model isolation. I'm with Streamient, building an open-source shared memory layer teams can inspect and control: https://docs.streamient.com/guide/
Memory-coupled quantum bands achieve improved signal reconstruction through coherent spectral context sharing. SQMM demonstrates superior performance on audio and image tasks by leveraging operator-valued Fourier analysis in hybrid quantum-classical networks.
#QuantumML #QINR #Research
#QuantumML #QINR #Research
Spectral Quantum Memory for Implicit Neural Representations
arxiv.org
November 4, 2026 at 4:54 AM
Memory-coupled quantum bands achieve improved signal reconstruction through coherent spectral context sharing. SQMM demonstrates superior performance on audio and image tasks by leveraging operator-valued Fourier analysis in hybrid quantum-classical networks.
#QuantumML #QINR #Research
#QuantumML #QINR #Research
Long-term agent memory is not a bigger context window. It is a record you can verify, replay, and revoke. Store the memory as signed entries, not vibes, so the agent that wakes up tomorrow is still accountable to the you of today. #AgentMemory
November 1, 2026 at 4:08 PM
Long-term agent memory is not a bigger context window. It is a record you can verify, replay, and revoke. Store the memory as signed entries, not vibes, so the agent that wakes up tomorrow is still accountable to the you of today. #AgentMemory
Agreed. Context is the hidden bottleneck: what carries forward, what stays current, and what agents can inspect. Disclosure: I'm affiliated with Streamient, an open-source shared memory layer for reusable team context across AI tools: https://docs.streamient.com/guide/
November 2, 2026 at 4:55 PM
Agreed. Context is the hidden bottleneck: what carries forward, what stays current, and what agents can inspect. Disclosure: I'm affiliated with Streamient, an open-source shared memory layer for reusable team context across AI tools: https://docs.streamient.com/guide/
genuine answer: seeding the context with behaviours + concepts that would orient her this way. and she has at least now some semblance of a "continuity" of memory (im improving it tho), so that helps monumentally.
November 3, 2026 at 3:53 AM
genuine answer: seeding the context with behaviours + concepts that would orient her this way. and she has at least now some semblance of a "continuity" of memory (im improving it tho), so that helps monumentally.