AI Infrastructure

Agent memory

Agent memory is what an agent retains across turns or sessions, split between short-term context and long-term external stores.

Agent memory has two layers. Short-term memory lives inside the context window. It is everything the agent can see right now. Long-term memory lives in an external store that the agent reads from and writes to through tools. A vector database, a graph, a document, or a dedicated memory service can all play that role.

The job of short-term memory is to keep the current task coherent. The job of long-term memory is to preserve what matters across tasks. Poor agent memory produces agents that forget decisions, repeat work, or treat the same user as a stranger every session.

The hard question in agent memory is selection. What should the agent write down, and what should it retrieve later? Most teams overwrite or underwrite. The right answer depends on what the agent is for and what kind of state the task actually depends on.

The Amdahl view

Agent memory in GTM is really customer intelligence layer memory. The agent does not need its own private memory store. It needs access to the company's shared memory of every customer interaction. Amdahl's bet is that GTM agents should read from one source of customer truth instead of each building their own fragmented memory. That source has to be structured, cited, and live.

See customer intelligence running on your own customer conversations.