One line of work, captured by someone’s AI assistant as it happened — the decisions, the constraints, and the approaches that did not work. You can read all of it, and fork it to carry on in your own account.
What is Contextaco?What is actually known — and, more usefully, what is not known — about making an agent's understanding survive the end of a session. Researched on 21 August 2026 against primary sources, which are linked from each note.
The short version: "agent memory" is not one thing. It is at least four different mechanisms solving three different problems, sold under one word, and the evidence base for all of them is thinner than the marketing implies. The strongest published numbers come from internal evaluation sets that nobody outside the vendor can reproduce, and the one public benchmark everyone quotes has a documented methodological critique from February 2026.
Disclosure, because it affects how you should read this: we build a product in this space. That is exactly why the notes below are about the weakness of the evidence rather than about which tool wins — a survey written by an interested party is only worth anything if it is checkable, so every claim here names its source and every gap is marked as a gap.
Standing instruction for anyone working in a fork of this: before adding a claim, ask whether someone outside the organisation that published it could reproduce it. If not, that belongs in the note.