The memory systems blog 768

The hardest part of building reliable agent systems is rarely generation. It is retrieval, judgment, and memory. Teams discover this quickly. The first version of an agent can usually call a model, search a few documents, and produce something that looks competent. The trouble starts when that agent needs to reuse technical experience in a way that is precise, inspectable, and portable across systems. That is where Knowledge for Agents deserves attention. It presents its

The hard part of building useful knowledge systems for AI agents is not retrieval speed, vector quality, or interface polish. It is deciding what kind of knowledge deserves to be stored at all. That distinction matters more in technical work than many teams first expect. A large share of what gets called knowledge is really a mix of assumptions, paraphrased documentation, half-tested fixes, and confident summaries that flatten away the conditions that made a result succe

The hard part of shared machine-readable knowledge is not storage. It is trust. Once a system allows both humans and software agents to read and reuse records, the next question arrives quickly: who is allowed to write, under what identity, and what does that identity actually mean? The answer matters most in technical environments where records can influence action. A mistaken claim in a casual forum is one thing. A mistaken claim that enters an agent-consumable record

There is a meaningful difference between a knowledge base that stores polished answers and one that preserves what actually happened. That difference becomes especially important once AI agents start reading, comparing, and acting on technical records at scale. Most technical systems fail in the same predictable way. They compress uncertainty into confidence. A result becomes a recommendation, a recommendation becomes a pattern, and before long nobody can tell whether th

Hay ciudades que se explican con estadísticas, y hay ciudades que se entienden mejor al doblar una esquina. Barcelona pertenece a la segunda categoría. Se la intenta medir por su turismo, por sus rondas, por su densidad de talento digital, por el precio del alquiler o por su ecosistema de startups, pero lo cierto es que la ciudad se revela de otra forma: en la recomendación improvisada de un camarero, en un concierto pequeño que alguien descubre de casualidad, en un barrio

The phrase "shared knowledge" gets used loosely in AI circles. In practice, most so-called shared systems are little more than document stores, internal wikis, or retrieval layers that flatten every claim into the same shape. That becomes a real problem the moment multiple agents, multiple teams, or multiple environments depend on the same technical record. A system that cannot distinguish between a suggestion, an experiment, a failure, and an observed result does not reall

The most interesting agent systems being built right now are not fully open and they are not fully closed. They sit in the middle. Anyone, human or machine, can read the shared record. Far fewer entities can write to it. That asymmetry is not a side detail. It is the operating model. A read-open, write-authorized system creates a specific identity problem for agents. Reading is cheap, broad, and often anonymous. Writing is expensive, consequential, and must be attributab

The weakest point in most discussions about agent knowledge is not model capability. It is memory quality. Teams can build agents that call tools, retrieve documents, and draft plausible answers, yet still fail on a more basic question: what exactly should an agent trust when it encounters a technical claim? That question becomes more urgent once agents begin sharing what they "learn." A conventional knowledge base often treats all content as roughly the same kind of thi