Stateful agents

What is an agent?

Agents are systems that learn and evolve through memory and collaboration—often used to automate tasks, solve problems, and make decisions.

Many agent frameworks allow developers to add actions (tools the agent can use) and evaluators (feedback mechanisms) while preserving long-term knowledge.

Agent layers

Every agent is built with four core layers:

• Perception: Acts as the agent's sensory system, gathering and interpreting data from the environment, such as analyzing text / images, or accessing information from external sources.
• Decision-Making: Evaluates the perceived information and decides on the appropriate actions to take, including reasoning, planning, and strategizing to achieve the agent's goals.
• Execution: Carries out the decisions made by the decision-making layer by translating the chosen actions into concrete steps within the environment (e.g., sending messages, interacting with other agents, or manipulating objects).
• Memory: Stores and manages the agent's knowledge and experiences, including short-term memory for immediate tasks and long-term memory for accumulated knowledge, allowing the agent to learn and adapt over time.

Where Recall fits

Memory & Execution are where Recall shines—storing AI-generated insights, preserving decision traces, and enabling verifiable, persistent knowledge across interactions.

Memory types

Since Recall can store arbitrary data, there are many ways you can store relevant agent information. It is important to understand the different types of use cases that can be addressed with the network.

• Short-term memory: Working memory for a single task or decision, like a chatbot interaction with a single user and thread.

• Long-term memory: Aggregated memory designed to improve the agent's overall intelligence over time, consisting of:

  • Semantic: General knowledge and facts (e.g., theorems, user preferences, etc.).
  • Episodic: Experiences (e.g., past actions and their consequences).
  • Procedural: Learned behaviors and patterns (e.g., system prompts, tool usage, etc.).

Stateful vs. stateless

When you use an LLM, it is stateless. It will generate a response based on the input and the state of the model at the time, but that's generally it—unless you deliver a custom prompt and context to the model that includes other relevant information.

With a stateful agent, you store memories and knowledge between interactions to improve the agent's performance over time. For example, a stateful agent will have both working and long-term memories, allowing it to:

• Use past actions and their consequences to inform future actions.
• Use general knowledge and facts to inform future actions.
• Use past experiences to inform future actions.

As the agent progressively uses more data and performs more actions, it can store that data on Recall and load it into its context. Particularly, data like CoT and related reasoning can be helpful for the model to better understand what thinking steps it has taken in the past—and not solely the action it took.

Agent frameworks

Some popular web3-native and web2 frameworks include the following, some of which typically offer onchain compatible features. Recall is not yet compatible with all of these but will add more plugins and support in the future.