Exxeta

Adaptive Chunking targets the RAG problem teams often debug too late: one splitter for every document. Put it in the ingestion step, let page, recursive, and LLM-regex chunkers compete, then score references, cohesion, surrounding context, block integrity, and size before indexing. No ground-truth QA set required. On 33 PDFs: retrieval completeness 67.7 vs 58.1, correctness 78.0 vs 70.1. #RAG #Chunking #contextEngineering #AIInfra #OpenSource

Memory frameworks compete on storage: Mem0 accumulates, MemoryBear decays, memvid keeps everything, Auto-Dream consolidates between sessions. SimpleMem competes on retrieval. An LLM generates a per-query plan over three indexes (semantic, lexical, symbolic), and EvolveMem v3 mutates that planner's config via evaluate-diagnose-propose-guard loops. 43.24% F1 on LoCoMo at ~550 tokens, 30x fewer than full context. The compression label is the wrong frame for what they shipped. #SimpleMem #AgentMemory #contextEngineering #LoCoMo #AIAgents

Tencent 's open-source memory plugin for OpenClaw cuts WideSearch tokens 61% and pushes pass rate from 33% to 50%. The mechanism: symbolic short-term memory. Verbose tool logs offload to disk; what stays in context is a Mermaid graph the LLM parses natively, with node_id pointers back to raw evidence. Cross-session memory layers L0 conversations into L1 atoms, L2 scenarios, L3 personas … PersonaMem jumps 48% to 76%. Local SQLite. 2.3k stars, MIT. #TencentDB #AgentMemory #OpenClaw #contextEngineering #AIInfra

DeerFlow 2.0 is ByteDance's open-source super-agent harness, 68k stars. Built-in skills ship for research, report generation, slide decks, web pages, image+video gen, loaded progressively. Sub-agents run in isolated sandboxes with filesystem and bash. Memory persists across sessions. Tasks come from a web UI, six IM channels including Slack and Telegram, or Claude Code via the `claude-to-deerflow` skill. Hand it a prompt, get back a finished artifact. #DeerFlow #AgentHarness #AIInfra #AgenticAI

TurboQuant landed as turbovec, the Rust implementation of Google 's ICLR 2026 vector quantization paper. A 10M-document RAG corpus that took 31GB of RAM or pushed you onto a managed vector DB now fits in 4GB on a laptop. No codebook training, no rebuilds when you add documents, no recurring bill. Faster than FAISS PQ on ARM. Reach for it when you run RAG locally and your corpus keeps growing. #TurboQuant #VectorSearch #Quantization #RAG #AIInfra

BMAD-METHOD hit 47k stars by shipping a planning loop that runs before the IDE loop touches code. Twelve agent personas (PM, Architect, Dev, UX, QA) produce PRD, architecture spec, and stories first; Claude Code or Cursor implements them after. Spec Kit and OpenSpec (99k and 48k stars) bet on artifacts you hand the agent; BMAD bets on personas that produce them. The planning step has its own harness now. #BMAD #SpecDrivenDevelopment #agenticCoding #AgentHarness

Hermes hit 140K stars in under 3 months. Qwen 3.6 27B claims 400B-class accuracy at 1/16 the parameters; Qwen 3.6 35B needs roughly 20GB of memory for local inference. DGX Spark gives agents 128GB unified memory for all-day runs. When inference moves into hardware capex, the moat shifts from token access to stack control. Local agents are no longer edge cases. #HermesAgent #Qwen #AIInfra #LocalInference #AgentFramework

Qwen3.6-27B, Unsloth AI Studio, Pi on your Mac: local coding agent, zero per-token cost, private repo context. Download the 16GB GGUF, point Studio at it, point Pi at Studio. Eight tokens/sec on Apple Silicon, reasoning built for agentic coding, not chat. The leverage is the harness. You own what loads, what runs each step, what stops the loop. #agenticCoding #localModels #costEngineering #AIInfra

macpow ranks per-process energy on Apple Silicon without sudo. It reads the kernel's per-process energy counters in userspace, plus direct hardware reads for CPU, GPU, ANE, and DRAM watts, real frequencies in MHz, and temperatures from M1 to M5. powermetrics had this behind root. asitop wrapped powermetrics. macpow goes around it. Your local model, your coding agent, your IDE … now you can see what they cost in watts. #macpow #AppleSilicon #LocalInference #AIInfra #Rust