love this frame. calls to mind the role of the hippocampus in human navigation (via place cells and grid cells), and how navigation is, in a…

love this frame. calls to mind the role of the hippocampus in human navigation (via place cells and grid cells), and how navigation is, in a sense, what makes agents agents vs plain old LLM calls in the first place. this "use case" of the hippocampus also sheds light, I think, on the selection problem @yoheinakajima highlights — the brain decides what to "select" via salience. my favorite framework was developed by caduff & timpf in 2008. they identified: 1. cognitive salience (top-down directed attention shaped by experience and knowledge, e.g. "that's my old apartment!" vs any other building) 2. perceptual salience (bottom-up attention pulled from contrasts in the environment, eg. "that weird-shaped intersection") 3. contextual salience (situational availability of attention e.g. "I was so focused on the basketball that I didn't notice the gorilla") ...all working together to determine what actually got stored in the hippocampus. ok, so agents — "A long-running memory that keeps everything at equal weight is no memory at all." I've seen several teams now try to solve "multiplayer" AI by making exactly this mistake: they create a shared repository of all their agent logs, so that technically everything that has ever happened is available to everybody and all agents. as yohei points out, it doesn't work. other memory systems are a little more judicious, but fail in other ways. 1. storage is probabilistic (a few months ago my agent stored a memory that I am "passionate about the right way to reheat a leftover burrito") and usually handled by the same loop responsible for task execution 2. recall is also probabilistic (yesterday my agent decided it made sense to remind me of my passion for reheated burritos when I was trying to fix my broken amplifier) 3. even when storage and recall happen reliably, actually doing the right thing with that information is faulty. there was a great study on #3 recently: https://arxiv.org/abs/2602.16313 the tl;dr is that markdown-based memory breaks when agents need to carry decisions across sessions. on 766 multi-session tasks, with ~57 action steps and long traces, the current SOTA for long-context, RAG, and memory-agent systems still had poor full-task success; the best average success was only about 23%, and some planning tasks were 0% across all methods. since claude.md is unstructured text, the agent has to notice, recall, and obey old decisions probabilistically, and this get increasingly difficult with scale. so, what is salient to an agent and how do we solve this problem? anyone who's used agents knows they are noisy token cannons. at @primitive_hq we believe the most salient information for an agent are the Decisions (and the rationale behind them) that you or your agent make in the course of your work, specifically those decisions that may influence future work. to extend the mice-in-a-maze model, if turning left immediately out of the gate results in cheese, then turning left is an important decision that should be made available to future agents (and teammates) so they don't waste time and tokens following dead-ends you've already identified—and yet the maze may change, and the system must be able to learn and flex with these changes. so the job, I believe, is for an outer-loop system to identify when consequential decisions are made, store them as an immutable causal graph (since yesterday's decisions inform today's decisions and today's decisions can't undo yesterday's decisions, only supersede them), and reliably enforce them on future agent sessions across the team without creating a rigid system that can't adapt. in other words: "Coordination was a property of the shared state and the order of the writes." very exciting to see experimentation like activegraph in this space along with many others — I strongly believe this is the right direction to explore

Source: Yohei Nakajima (X) | 2026-07-27