Agent Memory Needs a Control Plane
Executive Summary
The useful question for long-running agents is no longer simply whether they have memory. It is what they retain, when they retrieve it, and whether the current session can tell the difference between durable instructions, live project state, and historical record. A cluster of talks today converged on that point from practitioner, evaluation, and research directions: more context is not automatically more competence, and indiscriminate retrieval can make an agent slower or worse.
That complicates a familiar assumption in agent design—that a growing transcript or vector store naturally becomes an asset. The emerging view is more demanding: memory is a control problem, with explicit state, ranking, baselines, and review points.
What Happened
Nate B Jones offered the most operational version of the argument in a discussion of long-running coding agents. He separates stable instructions, a compact mutable record of the project’s current state, a map to relevant resources, and the historical record. The practical prescription is to make the first three legible to a fresh session while keeping history available but non-directive. In his telling, a rebuilt short plan stopped an agent from repeatedly reorganizing its own task list rather than progressing on the work. That is practitioner synthesis rather than a controlled study, but the design distinction is concrete: the entire past should not become the agent’s active prompt. Watch the talk
Stefania Druga of Sakana AI described a research-agent comparison that makes the same point with a sharper empirical claim. In a literature-review task where all relevant material already fit in context, added memory imposed cost without improving results. On a set of longer-horizon questions, she reported that a ranked decision-ledger retrieval scheme outperformed no memory, vector retrieval, and a gate that decided whether to use memory; it also used fewer tokens. Even an oracle retriever did not achieve perfect performance, because models can fail to use correct material well. The reported results need a published methodology and metrics before they should be generalized, but the evaluation lesson is sturdy: retrieval quality and model use of retrieval are separate problems. Watch the talk
Parth Asawa’s Continual Learning Bench framing pushes further. Instead of treating a benchmark score as evidence that a system learns, it compares a stateful system with the same system reset after every task; the difference in reward is the learning gain. His initial results reportedly put plain in-context learning on the best reward/cost and gain/cost frontiers, ahead of more elaborate context-management schemes, while acknowledging that the tasks may be too medium-horizon to settle the broader question. Watch the talk
Why It Matters
These are not merely variations on RAG. They split “memory” into at least four jobs: preserving commitments, representing current state, locating evidence, and adapting future behavior. A system may be good at one and bad at another. The day’s strongest contribution is a warning against merging them into one opaque store and calling the result learning.
Yu Su’s distinction between intelligence and expertise makes the stakes clear: a highly capable model can still be a novice in a particular organization or workflow if it cannot compress experience into reusable, situated knowledge. He argues that coding is unusually favorable because its work and feedback are structured, whereas enterprise and personal work consist of heterogeneous local contexts. Watch the talk
This reinforces the developing canon around agents: reliability is increasingly a harness and evaluation property, not just a model property. Bigger windows make more information available; they do not decide what deserves authority, what is stale, or what should be acted on now.
Workflow Implications
For builders, the immediate experiment is small. Create a short, explicitly owned current-state document for an agent: objective, accepted decisions, constraints, open risks, next verifiable actions, and pointers to evidence. Keep it separate from immutable instructions and from logs. At handoff or review checkpoints, ask a fresh session to validate and update that state against the actual repository or task environment.
Then evaluate the memory layer against a reset baseline. Measure task completion, regressions, token cost, and how often retrieval changes a correct decision into a wrong one. Test ranked decision recall separately from generic semantic search. If the agent is performing adequately with its active context, do not assume an additional store helps.