Overview
This series treats long-term agent memory as state management for evolving interactions, not a larger recall buffer. It traces how updates create current-state, historical, and aggregation questions; why action trajectories require causal links and exact operational details; where construction, storage, retrieval, prompt assembly, generation, and maintenance can lose state or add cost; and how freshness, workload, evidence fidelity, and tail latency shape architecture choice. MINTEVAL, AMA-Bench, and a systems-characterization study provide complementary measurements rather than one universal ranking: evaluated systems remain brittle under interference, causal dependencies, and changing workloads. The series ends with an update-fidelity contract covering identity, time, provenance, state change, causal dependency, and exactness, plus tests that separate construction, retrieval, reasoning, downstream behavior, and lifecycle cost.
Learning outcomes
- Update fidelity matters beyond isolated recall: current, historical, and aggregation questions need different evidence access.
- Agent trajectories require causal dependencies and exact operational details, not only semantically similar summaries.
- Construction, retrieval, freshness, and maintenance are separate failure and cost boundaries.
- No architecture is universally best; choose against workload semantics, update cadence, evidence needs, and resource limits.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Why can long-term agent memory retrieve a fact and still fail?
Because a plausible fact may have the wrong version, missing provenance, or no causal link to the needed action. Evidence can also be lost during construction, omitted by retrieval, or misused by the final model.
What should a memory system preserve when facts change?
It should preserve the entity identity, time or version, provenance, state change, and any causal relationship or exact detail the workload needs. Current, historical, and aggregation questions require different access to that evidence.
How do memory architectures trade accuracy, freshness, and cost?
They move work among construction, retrieval, and maintenance: richer structures may support history and relations at higher cost, while simple indexes stay cheaper and fresher but rely more on matching.
How should long-term agent memory be evaluated?
Evaluate current, historical, causal, and aggregation questions, then separate construction, retrieval, and reasoning failures. Also measure freshness, downstream action, latency tails, growth, and lifecycle cost.