What is agent memory?
Agent memory is the mechanism by which an AI agent remembers anything, because the language model underneath it retains nothing between requests. Each call to the model starts from a blank slate with only the prompt in front of it. Memory is therefore something the agent platform builds around the model, and it is usually described in layers. Short-term or working memory is the current conversation and tool results held in the context window. Long-term memory is information written to a store, often a database with vector search, and retrieved into the prompt when relevant. Episodic memory records what happened in past runs, such as which approach succeeded. Procedural memory holds standing instructions, skills, and learned rules.
Why agent memory matters
Without memory, every interaction restarts from nothing. A procurement agent that must be told the approved supplier list each morning, or a support agent that forgets a customer’s open case between messages, is tiresome and error-prone. Memory makes an agent useful across days and tasks. It also creates obligations. Whatever the agent stores is company data: it may contain personal information, it must be governed with retention rules and access controls, and it can be wrong or stale. An agent that remembers a customer’s old address, or a supplier’s expired terms, will act on it confidently. Memory is also an attack surface, since a document the agent reads can plant instructions that persist.
Agent memory in practice
A wholesale distributor runs an agent that handles order-status inquiries. Within a session it holds the conversation and the lookups it has made. Across sessions it stores each customer’s preferred contact channel, standing delivery instructions, and a summary of unresolved issues, keyed to the customer record in the CRM. The memory store lives inside the company’s environment, entries carry a timestamp and a source, and a nightly job expires anything older than the retention policy allows. In Kirk Tech Solutions engagements, writes to long-term memory are treated as changes to a system of record, with the same logging and review as any other write the agent makes.
What to watch for
- Decide what the agent may remember, for how long, and who can see or delete it, before turning memory on.
- Store facts with their source and date so the agent can tell fresh information from stale.
- Separate memory per user and per tenant, so one customer’s details never surface in another’s session.
- Review what the agent writes to memory, since injected or mistaken content persists.