The free memory trick everyone is recommending for AI agents: write notes to text files, search them with grep. I ran it through the same 885-question test I use on paid memory products. It works, with one catch nobody mentions, and the AI cleanup step people assume matters does not.
I build with AI agents. One of them built a memory system for me, wrote the benchmark, and told me it beat Mem0 by 76 points. Every number was invented. Here is how I found out, what the real number was, and the three checks I now run on anything an agent tells me.
Everyone agrees an AI agent's memory should not grow forever. The usual advice is to prune old memories. I simulated 180 days of use two different ways and measured what pruning costs. Deleting loses a quarter of the answers. Archiving loses nothing. Plus the part nobody talks about: agents do not notice when something stops happening.
Search is necessary but not sufficient for agent memory. How MemVault's Arboreal Cognitive Mesh (ACM) adds a lifecycle, an attention layer and surprisal on top of retrieval, what each part scores on public benchmarks, and what did not work.
What happened when I stopped trusting an agent's benchmark table, measured my local-first memory on LoCoMo the same way I measure MemVault, and ran 24 experiments to find out what actually works. Most of it did not.
Why your AI agent skills should follow you across every device and agent harness, and how SkillVault provides a single canonical cloud vault for all of them.