liveBook · Manning

Memory

As explained in AI Governance

Agent memory is information retained by an agent for use in later interactions or tasks.

An agent that remembers past interactions can resume work and adapt its strategy, but it can also retain personal data across sessions, users, and services. Agent memory is information retained for later use: it can include completed actions, outstanding tasks, intermediate results, and durable facts. section 2.6.2 section 3.8

Memory can serve different purposes:

  • Working memory holds temporary scratchpad information for the current task, such as sub-goals, partial outputs, and tool results. section 2.6.2
  • Task memory records actions and outcomes so work can resume after interruptions and support audits. section 2.6.2
  • Long-term memory stores searchable facts, preferences, and artifacts. Multi-agent systems may also use a shared store for collaboration. section 2.6.2

Patient monitoring

A patient-monitoring agent can retain baseline health metrics, symptoms, medication adherence, and prior alerts. It can compare new readings with that history to spot patterns, such as a gradual rise in resting heart rate alongside increased inhaler use, before deciding whether to seek more information or escalate. section 2.7

Persistent memory needs governance because an agent may decide what to store, when to surface it, and whom to share it with. Define retention periods, deletion rights, data subject access procedures, and tenant-scoped access controls. A conventional centralized retention schedule may not account for information an agent stores across multiple underlying stores. section 2.6.2 section 3.8

Design deletion and isolation before building the store. For patient data, tag each record with its patient and care episode, and partition storage so a search for one patient cannot reach another’s records. Deletion should cover the whole relevant partition, including summaries, raw exchanges, and embeddings. section 3.8

Trace retention beyond the initial agent. A doctor’s appointment task may send health data to calendar, transit, and insurance services, each of which may retain it independently. Set retention timelines and arrange deletion across the services touched; automated scripts can purge aged caches, such as chatbot transcripts older than 60 days. section 5.6.3

How AI Governance covers it

How deep it goes
Distinguishes memory types and gives retention, partitioning, and deletion guidance without specifying a memory-store implementation.
Examples
Illustrates memory with patient-monitoring history and health data distributed during appointment booking.
What you can do after
Evaluate memory designs for retention, patient isolation, and deletion across stored representations and downstream services.
How it connects
Connects persistent memory to multi-agent systems, access control, data retention, and the right to erasure.

Also explained in

Compare how 26 books explain memory →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.