liveBook · Manning

Memory

As explained in Designing AI Agents

An agent capability for retaining or retrieving information across or within tasks.

Memory is the policy for what information survives a boundary and under what authority it may return to a later decision. It owns shared-state persistence and retrieval; a separate memory service is optional. section 4.9 section 2.2.1

Other cognitive modules use memory selectively: perception can retrieve prior context, reasoning can retrieve earlier evidence, and reflection can propose lessons. Raw evidence and candidate lessons may be preserved, but only admitted lessons may influence live or later work. Recency and frequency can guide retrieval; they do not establish authority. section 2.2.1 section 4.9

Select memory patterns from observed failures rather than treating them as a mandatory pipeline:

  • Hierarchical retention separates scope, function, access, and lifetime. section 4.9
  • A RAG pipeline governs source compilation and assembly of decision evidence. Retrieval may be exact, lexical, semantic, structural, relational, or temporal, depending on the corpus. section 4.9
  • Progress tracking preserves the goal, acceptance condition, artifact evidence, authoritative versions, pending decisions, side-effect receipts, and handoff conditions needed to resume safely. section 4.9
  • Failure journals distinguish raw failure evidence, candidate diagnoses, and published lessons; admission and replay are separate responsibilities. section 4.9

Delegated handoff

A delegated worker may share a repository without inheriting its parent’s conversation or tool history. Memory supplies the versioned findings and source references needed for the handoff; collaboration determines the receiving role, task, and acceptance of the integrated result. section 4.7

Evaluate memory against downstream decisions within a defined workload. Useful measures include retrieval utility, repeated mistakes, stale reads, duplicated effects, harmful transfer, latency, and token cost. section 4.9

How Designing AI Agents covers it

How deep it goes
Defines retention authority and summarizes four optional patterns, their responsibilities, retrieval choices, and evaluation measures.
Examples
Uses a delegated worker handoff and an authoritative team rule in a controlled rollout to illustrate retained context.
What you can do after
Evaluate memory designs against downstream decision quality, failure recurrence, latency, and token cost.
How it connects
Connects memory with perception, reasoning, reflection, collaboration, and retrieval-augmented generation.

Also explained in

Compare how 26 books explain memory →

Related concepts

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