Open Source Releases

lucid-yolo 0.1.0.dev3 — From-scratch PyTorch Lightning implementation of YOLO26. Covers detection, instance segmentation, and oriented detection in one codebase. Solid reference if you’re dissecting modern YOLO internals or need a clean starting point for customization.

m3-memory 2026.8.9.0 — Local-first agent memory hitting 99.2% on LongMemEval-S. Hybrid search (FTS5 + vector + MMR), one-command MCP plugins for Claude/Gemini, and a setup wizard that doesn’t suck. Finally, memory infrastructure that treats local-first as a feature, not an afterthought.

agent-runtime-governance 0.9.1 — Runtime guardrails for production agents: exact-call approval, idempotent execution, signed audit trails. The “UNKNOWN outcome” recording is a nice touch — admits when something went sideways instead of hallucinating success. 🛡️

context-gc 0.2.1 — Deterministic context compaction middleware with receipt preservation. Compresses agent history while keeping an audit trail of what got tossed. Predictable window management without the “trust me bro” opacity.

lorekeep 0.23.1 — Temporal knowledge graph with MCP integration. Time-aware knowledge representation for agents that need to remember when facts were true, not just that they were. Useful for anything with evolving state.

textgraph-kg 3.3.0 — Local-first KG construction with byte-level provenance on every claim. Zero-LLM-by-default, deterministic, reproducible. If you need to explain why your pipeline extracted a relation, this gives you the receipts down to the byte offset. 📄

Research Worth Reading

Universal Pathologies, Conditional Consequences — Triple-robustness analysis of GraphRAG vs vector RAG across embedder, corpus, and query type. Finally someone systematically mapped where citation precision dies instead of just claiming “GraphRAG is better.” Actionable architectural guidance included.

SemiAdapt-Instruct — Discovers latent instruction domains, trains per-domain LoRA adapters in parallel, composes at inference. Extends fine-tuned models to new domains without full retrain. The “extensible instruction tuning” framing is marketing, but the modular adapter composition is genuinely useful for evolving deployments.

Simulator-Grounded LLMs for Industrial Causal Reasoning — Grounds frozen Qwen2.5-32B in a wastewater treatment simulator via three approaches: tool-use, causal graph injection, and plant-portable retrieval. Real industrial causal reasoning benchmark — rare to see this level of domain specificity.

Where Privacy Risk Lives in English-Source Multilingual RAG — Stage-decomposed privacy audit across five query languages. Tests whether non-English queries actually increase PII leakage (spoiler: it’s complicated). Evaluates a two-stage defense with judge + regex filter. Practical if you’re shipping multilingual RAG.

RIG-RoPE — Enhanced rotary positional encoding with relation/instance gating and duration-aware temporal coordinates. Addresses static position assignment limits in multimodal RoPE. Niche but relevant if you’re pushing temporal reasoning in multimodal contexts.

Conditional Cognitive Biases in LLMs — Three-condition framework disentangling biased user turns from semantic content effects. Benchmarks 24 biases across multi-turn interactions. Turns out user framing really modulates in-context reasoning — instruction tuning doesn’t immunize models.

AI Dev Tools

prime-agent — RLM agent for autonomous coding with self-generated feedback loops. Iteratively improves on long-running workflows. The “self-improving” claim is bold; the reinforcement learning from machine feedback approach is worth watching if you’re automating software engineering tasks.

semantica — Graph-native infrastructure for accountable AI: context, provenance, data lineage, model decisions tracked in knowledge graphs. Enables auditable RAG and multi-agent coordination with explicit state. “Accountable AI” is a buzzword magnet, but the graph-based lineage tracking solves real debugging pain.

swarm-forge — Lightweight multi-agent coordination from Uncle Bob. Clean architecture patterns for agent swarms: message passing, shared memory, failure recovery. Minimal by design — refreshing in a space where every framework tries to be a platform.

Last30Days Skill — Agent skill that researches across Reddit, X, YouTube, HN, Polymarket, web — then synthesizes grounded summaries with recency filtering. Multi-platform aggregation for up-to-date retrieval. Handy for agents that need current context, not just training cutoff knowledge.

Today’s Synthesis

If you’re building agents that run for days instead of minutes, the memory stack is where reliability lives or dies. m3-memory gives you local-first storage with hybrid search and MCP plugins that actually work; context-gc adds deterministic compaction with receipt preservation so you never lose why a decision was made; lorekeep layers temporal knowledge graphs on top so agents track when facts were true, not just that they were. Wire them together: m3-memory as the persistence layer, context-gc as the window manager that compresses history without silently dropping evidence, lorekeep as the temporal index for evolving state. The result is an agent memory system that’s auditable, local-first, and survives context-window pressure without hallucinating its own past. Start with m3-memory’s setup wizard, add context-gc middleware on the message bus, then introduce lorekeep when you need time-aware queries. This isn’t theoretical — each piece is pip-installable today.