@Mr_Buzzoni - Jev Agent Memory Cuts

Open original

@Mr_Buzzoni on X This is a short X thread from Mr_Buzzoni describing Jev, a five-layer memory architecture for AI agents where each layer makes one typed decision, such as what gets a slot in the context window or what gets deleted. The author claims it cut one setup from 26,000 tokens per query to 1,800, with Claude building the memory and Jev deciding over it.

  • Agents that reread the same files each morning pay for the same work repeatedly; Jev is pitched as a fix for that cost.
  • The design splits memory into five layers (working, episodic, semantic, procedural, forgetting), each answering one typed question with an answer and a probability in about 100 ms.
  • The headline result is a reduction from 26,000 to 1,800 tokens per query, which the author describes as over 90% fewer tokens on every query.

The core claim is that an AI agent keeps re-reading the same 12 files every morning, re-deriving the same conclusions, and paying for those tokens each time. The author's fix is a memory layer he calls Jev, which sits on top of Claude and makes the memory decisions. In one setup he reports cutting a query from about 26,000 tokens to 1,800, a drop of more than 90 percent. The pitch is less about storing more context and more about deciding what context is worth loading in the first place. The architecture splits memory into five layers, each with a single decision attached. Working memory scores what earns a slot in the context window. Episodic memory picks which past events get replayed. Semantic memory checks a fact before it gets written. Procedural memory chooses which approach gets promoted into a reusable skill. Forgetting memory rules on what gets deleted. Each of these is framed as a typed question put to Jev, which returns an answer along with a probability, reportedly in about 100 milliseconds. The division of labor is the point: Claude builds the memory structure, and Jev makes the calls over it. The article is a short promotional thread, so the numbers are the author's own and come without a published benchmark, dataset, or description of the task being measured. Before adopting the approach you'd want to know what the 26,000-token baseline included, whether the 1,800-token result holds across different query types, and how often Jev's probability scores turn out wrong in ways that drop something the agent later needs. The underlying idea, that memory should be gated rather than dumped wholesale into context, is sound and worth borrowing regardless of whether this specific tool is the right vehicle.

Ask about this article