Agent memory

Models do not remember anything between calls; every call starts from the context it is given. Agent memory is the engineering that gives an agent continuity. Short-term or working memory is the state of the current task: what has been retrieved, which tools were called, what the plan was. Session memory is the conversation so far, often summarised as it grows. Long-term memory is what persists across sessions: a customer's preferences, past issues, facts the agent confirmed, stored in a database or vector store and retrieved when relevant.

Good memory design decides what to store, for how long, and who may read it. Storing everything is a privacy liability and degrades retrieval; storing nothing forces users to repeat themselves. Most production agents store structured facts with a source and a timestamp, retrieve a small relevant set per task, and let users see and correct what is remembered.

Agent memory is not the model's training data and not fine-tuning; it is data managed by the application. It is also not the context window, which is the temporary space a single call can see; memory is what decides which pieces get placed into that space. Memory holding personal data is subject to consent, retention and access rules like any other store.

• Product manager: memory is what makes an assistant feel like it knows the customer, which drives adoption and retention.
• Compliance officer: remembered facts about people are personal data and need consent, retention limits and deletion on request.
• CTO / Head of Engineering: memory design affects cost, latency and accuracy; retrieving too much fills the context window with noise.
• Support manager: an agent that recalls a customer's previous issue resolves faster and escalates with better context.

An agent that forgets everything between calls cannot complete multi-step tasks or serve the same customer twice without starting over. Memory exists to give agents continuity within a task and across time, so they can act on what they have already learned. The failure it prevents is the assistant that asks for the order number three times. The trade-off is that memory is state, and state brings cost, staleness, privacy obligations and the risk of acting on outdated or wrong facts. Memory must be scoped per user and tenant, given a retention policy and made visible and correctable.

• A SaaS copilot remembering a user's preferred report format and last-used filters across sessions.
• A banking assistant recalling an open dispute so the customer does not have to re-explain on the next contact.
• A patient-facing scheduling agent remembering a preferred clinic and language, stored under consent.
• A D2C WhatsApp assistant recalling sizes and past orders to make reorders one message long.
• A tutoring agent tracking which concepts a learner has mastered to choose the next exercise.
• A dispatch agent holding the state of a multi-day exception across several customer contacts.

Technique / practiceAn engineering technique combining state management, retrieval and data governance. Eazyware designs memory per agent under multi-agent systems and retrieval and knowledge engineering, scoped per tenant and user, with retention and deletion built in.

Eazyware service that covers it: Multi-Agent Systems & Workflow Orchestration. Starting prices are on the pricing page.