AI digest: Memory, money, and model madness
Tencent drops a proper memory system for agents, DeepSeek keeps undercutting everyone, and Anthropic's bug-finder is working too well.
Three releases this week that actually move the needle forward, plus some classic Silicon Valley drama.
Tencent builds memory that might actually work
Tencent open-sourced TencentDB Agent Memory, a 4-tier memory system for AI agents that runs entirely locally. The clever bit is how it compresses verbose tool logs into compact Mermaid task canvases for short-term memory, then builds a pyramid from conversations to personas for long-term storage. Ships as both an OpenClaw plugin and Docker image with hybrid BM25 + vector search. Finally, someone’s thinking seriously about agent memory instead of just bolting on more context windows.
DeepSeek makes stupid-cheap pricing permanent
DeepSeek made its 75% discount permanent, pricing output tokens 34x cheaper than GPT-5.5 at $0.435 per million input tokens. This is either a loss leader to grab market share or Chinese compute economics are genuinely different. Either way, it’s going to squeeze Western providers hard, especially for token-heavy agentic workflows where cost actually matters.
Anthropic’s bug finder works too well
Anthropic’s Claude Mythos Preview found over 10,000 critical vulnerabilities faster than developers can patch them. The model is part of Project Glasswing with 50 partners, and Anthropic is warning this creates a security nightmare. Classic AI problem, this one. The capability arrives before we’ve figured out how to handle the consequences. Someone needs to start thinking about responsible disclosure at AI speed.
Nous Research skips the SAE training dance
Nous Research released Contrastive Neuron Attribution, a method for steering LLM behaviour by identifying and ablating sparse MLP circuits without training sparse autoencoders or modifying weights. If this actually works reliably, it’s a much cleaner approach to model steering than the current mess of techniques everyone’s using.