Modern AI coders resemble high-skilled amnesiacs forced to re-read an internal wiki every time they move a finger. In his paper "Delivery, Not Storage," Swapnanil Saha diagnoses a systemic bottleneck: architectures reliant on conscious retrieval (RAG). We are forcing agents to decide for themselves what to commit to memory and when to recall it. The result is a catastrophic waste of resources and frequent operational failures.
Human expertise works differently. A seasoned developer doesn't search the documentation for every project-specific quirk; relevant "gotchas" surface in the mind involuntarily as a reaction to the task's context. Saha argues that for autonomous systems, memory must evolve from a tool the agent chooses to use into an inherent "harness property" of the control loop.
The Failure of Volitional Memory Tools
The idea of giving agents tools to manage their own memory has failed in practice. Research shows that prospective memory—the ability to act on hidden constraints when a trigger appears—is orders of magnitude harder for models than standard retrospective search. In a controlled test, an agent equipped with all necessary knowledge and memory tools failed to make a single archive request across 114 iterations. Even when the tools are at its fingertips, the model simply ignores them.
A reliable memory channel for an agent is one it doesn't have to think about.
The gap between data storage and delivery exposes a fundamental flaw: we still treat memory as content to be shoved into a context window, rather than a deterministic mechanism of the environment.
Trigger Injection vs. Context Degradation
The "Cue-Anchored" model proposed by Saha transforms memories into first-class objects with rigid activation conditions: file paths, symbols, events, or timestamps. In this framework, it is the execution harness, not the agent, that analyzes the situation and "injects" relevant facts at the precise moment they are needed. The efficiency was proven via a brutal context-compression stress test: ten critical facts held in simple chat history evaporated during the very first summarization. Using Cue-Anchored delivery, that same data survived 138 cycles without a single loss.
Product is delivery, not storage.
By the end of the experiment, the situation turned absurd: an agent deprived of proper memory began manually "grepping" its own session files from the disk, attempting to DIY the very operational layer the framework should have provided. This is a clear signal to the market: it is time to stop bloating context windows and start moving memory from the content plane to the control plane. For engineers, this means shifting from "smart" RAG to rigid, deterministic structures that feed context based on file paths and symbols. An agent should be a passenger of its own expertise, not a failing librarian.