AI Coding Live News — Real-Time 𝕏/Twitter Updates & Breakthroughs

Tencent Hunyuan@TencentHunyuan·Collected 18d ago

👀60

🔥86

Tencent Hunyuan open-sources Hy Memory 2.0: ~3B-active memory writer averages 55.49 across four harnesses (vendor-reported, edging Opus 5), plus MemHub local memory for Claude Code, Cursor and Codex

On Oct 16 Tencent Hunyuan open-sourced Hy Memory 2.0 in three layers: the Apache-2.0 memory-writer model Hy-Memory-30B-A3B (~3B active, HF weights plus arXiv report), a Python SDK covering Chat, Working and Coding scenes (PyPI hy-memory 2.0.x), and local-first MemHub (CLI + macOS, preview). Per Tencent, swapping it in as the writer across Mem0, MemOS, Hy-Memory and held-out EverMemOS yields a 55.49 average vs 54.62 for Claude Opus 5, measured on fixed 20% subsamples; it still trails Opus 5 on several individual metrics.

Key Takeaways

Vendor-reported: 55.49 four-harness average vs 54.62 for Claude Opus 5 and 53.64 for DeepSeek V4.1 Flash
Held-out EverMemOS: 56.89 vs Opus 5 at 54.04; still trails on Mem0 LoCoMo (59.54 vs 69.47)
~3B active, Apache-2.0, tested on vLLM v0.31.0 with an OpenAI-compatible API
SDK installs via pip install hy-memory with automatic Chat/Working/Coding routing
MemHub runs on Linux and macOS (preview); Windows unsupported

Read full story

Linked Models & Agents:

⚡ Hy-Memory-30B-A3B⚡ Claude Opus 5 ⚡ DeepSeek-V4.1-Flash 🤖 Claude Code (Anthropic)🤖 cursor 🤖 codex 🤖 opencode 🤖 qwen-code 🌐 Tencent (腾讯混元)🌐 Anthropic 🌐 DeepSeek

#Hy Memory#MemHub#Mem0#MemOS

Hugging Face Daily Papers@HuggingFace·Collected 28d ago

👀58

🔥99

Engram: Pre-Registered Agent Memory Trial Shows Raw Turn Selection Matches LLM Fact Extraction at 3,061x Lower Cost

Conversational memory for LLM agents traditionally divides into two opposing camps: expensive pipeline extraction that distills dialogues into structured factual memory stores versus raw conversational turn retrieval. Literature has reported contradictory findings. Researchers conducted a pre-registered, double-blind controlled study on held-out LoCoMo and LongMemEval benchmarks (arXiv:2609.34227). The investigation reveals that under constrained context budgets, selecting raw conversational turns via a lightweight typed decision model (Jev) is statistically non-inferior to full LLM fact extraction pipelines (95% one-sided bound -3.0 points vs. -5 margin)—while reducing write-time computational costs by a massive 3,061 times. Furthermore, the authors resolve literature contradictions by proving that reranking gains scale inversely with token budgets, while aggressive reranking degrades an agent's correct abstention reliability. All study protocols, data, and code are open-sourced at github.com/ris3abh/Engram.

Key Takeaways

Presents Engram, the first pre-registered double-blind study testing raw-turn selection against LLM fact extraction for agent memory
Shows raw turns selected via typed decision model Jev are statistically non-inferior to LLM fact extraction at 3,061x lower write cost
Proves reranking gains decay sharply as token budgets expand, while aggressive reranking degrades an agent's correct abstention rate

Read full story

Linked Models & Agents:

⚡ Engram⚡ Jev⚡ LoCoMo⚡ LongMemEval🤖 Agent-Memory 🤖 Typed-Decision-Model 🤖 Turn-Selector 🌐 ris3abh 🌐 Hugging Face

#Engram#Agent Memory#Fact Extraction#RAG#Conversational Memory#Benchmark

ByteDance DeerFlow@bytedance·Collected 38d ago

👀123

🔥95

DeerFlow 2.1.0: 772-PR milestone — verifiable subagent delegation, pluggable memory, five remote sandboxes

ByteDance’s DeerFlow shipped v2.1.0 on 2026-09-24, closing the milestone with 772 merged PRs (~83k GitHub stars). Focus: trust and ops—verifiable subagent delegation, optional durable batch_task, pluggable memory (DeerMem / OpenViking / mem0 HTTP), plus E2B, BoxLite micro-VM, Tenki and OpenSandbox providers. Breakings: unconditional X-Trace-Id, reserved /mnt/skills, memory config under memory.backend_config, Docker bind defaults to 127.0.0.1.

Key Takeaways

Shipped: v2.1.0 — 772 PRs; site deerflow.tech
Verifiable subagent delegation + tamper-evident tool stamps; optional durable batch_task
Pluggable memory.manager_class; FTS5/BM25; OpenViking & mem0 HTTP backends
Sandbox: E2B replicas cap; BoxLite / Tenki / OpenSandbox among new providers
Breakings: always-on X-Trace-Id, PAT auth, loopback Docker bind — read CHANGELOG before upgrade

Read full story

Linked Models & Agents:

⚡ Doubao Seed 2.0 Code⚡ DeepSeek-v3.2⚡ Kimi-2.5🤖 DeerFlow 🤖 SuperAgent 🌐 ByteDance 🌐 Volcengine

#DeerFlow#ByteDance#SuperAgent#Sandbox#Memory#MCP#Subagents#PAT#E2B

Mem0@mem0ai·Collected 39d ago

👀70

🔥90

Mem0 Python SDK 2.2.1: add() stops fake ADD successes; Bedrock Claude reasoning + Turbopuffer filter/score fixes

Mem0 Python SDK v2.2.1 (2026-09-25, PyPI mem0ai 2.2.1) makes Memory/AsyncMemory.add() report only vector-store inserts that actually succeeded—raising VectorStoreError when none insert—restores context-manager close semantics, fixes Bedrock Anthropic Converse to read the first text-bearing content block (after optional reasoningContent), applies full Turbopuffer filter operators, and remaps euclidean(_squared)/S3 Vectors scores to 1/(1+distance). Same-day Node SDK ts-v3.3.1 stops injecting OpenAI default baseURL/model into non-OpenAI providers.

Key Takeaways

Shipped: Python v2.2.1 / PyPI mem0ai 2.2.1; Node ts-v3.3.1 same day
Correctness: add() returns only inserted memories; raises VectorStoreError if none insert
Bedrock Anthropic: read first text-bearing Converse block after optional reasoningContent
Turbopuffer filters: full eq/ne/gt/gte/lt/lte/in/nin; unsupported ops raise ValueError
Scores: euclidean_squared and S3 Vectors euclidean map via 1/(1+distance) to avoid negatives/collapse-to-zero

Read full story

Linked Models & Agents:

🌐 mem0 🌐 Amazon AWS (Nova)🌐 Anthropic

#Mem0#agent memory#VectorStoreError#AWS Bedrock#Turbopuffer#S3 Vectors#Python SDK

mem0@mem0ai·Collected 65d ago

👀40

🔥88

View on 𝕏 ↗

Mem0 Integrates as Core Long-Term Memory Provider for Nous Research Hermes Agents

Mem0 integrated as the dedicated long-term memory layer for Nous Hermes agents, removing the legacy 2,200-character in-prompt memory ceiling with cloud, Docker, and in-process deployment options.

Key Takeaways

Overcomes Hermes's 2,200-character memory limit to enable continuous cumulative learning across sessions;
Flexible three-tier deployment architecture spanning managed cloud, self-hosted Docker, and embedded OSS;
Preserves intuitive markdown memory conventions while offloading vector indexing and semantic retrieval.

Read full story

Linked Models & Agents:

⚡ Nous Hermes 3