Token-maxxing is dead. Agentic memory is what comes next.
Presented by MongoDB We have been building databases as an industry for roughly 60 years . We have been building AI agents, in the form most people mean when they say the word today, for about 18 months. Sit with that ratio for a second, because it explains almost everything about the state of agentic development right now. Six decades versus a year and a half. We are not in the middle of this…
The industry has been building databases for around 60 years, while AI agents have been developed for about 18 months. Despite this significant difference in time, the state of agentic development is still in its infancy. There isn't yet a standardized architecture for agents, similar to how there isn't a typical "LAMP stack" for databases.
One early insight came from a period of interest in token-maxxing, where the focus was on minimizing token consumption. However, this approach quickly faced backlash as it measured activity rather than results. The underlying lesson from token-maxxing was that the context window, or the amount of information a model can handle at once, is the limited resource in agentic development.
Instead of trying to pack more data into every prompt, the focus should be on selecting what truly belongs in the context window. A good agentic memory system should store and retrieve information generated by the model in previous sessions, apply role-based access control to that data, and use semantic search to find relevant prior content.
This is a shift from the exact-match lookup of traditional databases. The architecture that emerges from these elements often combines a leaner, more cost-effective model with a powerful memory system. The leaner model acts as a judge, determining whether a generated answer from the memory system is sufficient or if it needs to generate a new response.
This approach can significantly reduce costs over time, as the expensive model isn't used for every query. In the future, agentic memory systems will not be a simple, flat storage of information, but will have different types of memories, similar to human memory. These types could include taxonomy for organization-specific terms, procedural memory for task lists and sequences, and more.
Importantly, not all generated content will be equally valuable, and humans will likely play a role in curation, promoting high-value memories and pruning less useful ones. This process, reminiscent of managing knowledge bases and documentation, will be crucial for turning agentic memory into a valuable asset rather than a collection of unused data.
Written by urgent.news from VentureBeat's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
