The Quiet Revolution in AI Memory

This week Anthropic dropped Claude Sonnet 4.5. The coding improvements are real, but the for me the most interesting feature isn’t about code at all - it’s about memory.

Claude Code now has a memory tool that can persist files to disk as markdown, extending its hold on the context window. If you’ve read the MemGPT paper, you’ll know why that matters: tool use plus memory is what turns a model call into an agent.

Anthropic has long pushed context compaction - pruning and compressing information so LLMs don’t drown in irrelevant detail. Even with expanded windows, too much noise degrades performance. Compaction is compression, and compression is a form of generalisation - one of intelligence’s core tricks. Storing memory, then, isn’t about hoarding data; it’s about saving it intelligently.

Anthropic isn’t alone. ChatGPT now remembers. Google is publishing on memory management. Today it looks like text files. But human memory isn’t a flat log - it’s a web of episodes, relationships, and abstractions. AI memory will have to move the same way.

In maths, that structure is a graph. In computing, a knowledge graph. It’s hard to imagine long-term AI memory without one. Hybrid systems that mix graphs with text and vectors already hint at the richness needed for durable, contextual recall.

Soon, this will be baked into the models themselves. Which means organisations need to prepare now - by structuring their own information the same way, so it slots seamlessly into the machine’s memory. That doesn’t mean a chaotic tangle of links. It means ontologies: frameworks that give connections meaning.

And it means open standards. The last thing we need is each foundation model building a walled garden around its memory. The semantic web - designed from the start for an agentic world, for people not just profit - already points the way.

⭕ MemGPT Paper: https://arxiv.org/abs/2310.08560

⭕ Context Rot: https://www.knowledge-graph-guys.com/blog/context-rot

⭕ Super Agents: https://www.knowledge-graph-guys.com/blog/walmarts-superagents-why-semantics-and-knowledge-graphs-are-the-real-foundation

⭕ Semantic Agents: https://www.knowledge-graph-guys.com/blog/simulations-vs-models

⭕ Semantic Compression: https://www.knowledge-graph-guys.com/blog/semantic-compression

⭕ The Great Compression: https://www.knowledge-graph-guys.com/blog/the-great-compression

⭕ Personal Agents: https://lnkd.in/eD5xy4Xr