The context window—the maximum amount of information a model can read in one go—ranges from tens of thousands to millions of tokens, but a longer history does not mean better recall. We believe Agent memory is essentially a context management engineering problem: once model parameters are trained and fixed, “remembering” depends either on content remaining in the current context or on external storage that can be read back later.

What This Is

The article divides long-trajectory management into three approaches. Trimming means putting less content into the window. Compression replaces older conversations with summaries that preserve requirements, progress, conclusions, and outstanding tasks. Externalization writes information to files while leaving only an index in the context.

All three can ease capacity and cost pressure, and all three carry costs. Trimming may discard information needed later. Compression requires an additional call and may lose details. Externalization depends on the Agent proactively reading the files afterward. The article also warns against waiting for the window to fail: once content approaches the limit, “context rot”—the gradual neglect of key points—may already have begun.

Industry View

Our judgment is that differentiation in Agent products is shifting from tool-use capability to the ability to maintain task state over the long term. Enterprises that merely enlarge the model context window may not solve reliability, cost, or latency problems. The memory layer will therefore gradually become independent infrastructure.

The risks are equally clear. If a summary omits a critical constraint, subsequent execution may continue in the wrong direction. If files lack structure, versioning, and audit mechanisms, “memory” can become a repository of outdated information. Compression may also invalidate caches and trigger additional charges. Long-term memory must therefore pursue more than storing as much as possible: it must remain traceable and updateable, with the ability to return to the original material for verification.

Impact on Regular People

For enterprise IT: When deploying Agents, budgets should shift from model calls alone toward state management, including context budgets, summarization strategies, external storage, and log tracing.

For individual professionals: Assistants that can consistently retain project context, confirmed conclusions, and outstanding tasks will be more useful. Still, important decisions should remain in documents or task systems so that errors from summaries do not accumulate over time.

For the consumer market: As personal assistants begin preserving preferences across sessions, users will gain a more continuous experience while paying greater attention to what information is stored, whether it can be deleted, and how incorrect memories can be corrected.