AI Agent Architecture (Paperback)

Paperback. The context window is a cache, not a store - and an agent that keeps its history in the prompt has confused the two.Atlas, Meridian Supply Co.'s customer-operations agent, learned to take many steps safely in Book 2. It still began every run empty. When Kestrel Labs asked any news on my earlier request?, Atlas answered a request it had no way of knowing about; when Lena put the previous replies in the prompt to fix it, Northgate Tooling's history overflowed the window in the second week - and before it overflowed, the model had quietly lost the one fact in the middle that the customer cared about. This book builds the memory subsystem that a many-step agent needs, one store at a time, each justified by an incident and proved by a test.Every chapter is a lab on the companion repository - one dependency, fully offline with a scripted mock, every listing printed from a verified line range, every command paired with its expected output. Five moves per chapter: Run the demo, Read the listing, Break it with the chapter's planted failure, Fix it with the design move, Prove it with a test that turns green. You will: Every chapter carries a research lineage (Atkinson-Shiffrin and Tulving, Lost in the Middle, case-based reasoning, BM25, RAG and the Cranfield tradition, GraphRAG, consolidation, data-protection retention, optimistic concurrency), a five-item failure catalog, an applied deep-dive, and exercises in three tiers. Ten figures, a hundred and thirty-three tests, twenty-six decision records, one running system that grew from Book 2's graph without removing a node of it.Who it's for: engineers designing what agents remember and know - anyone who has bolted a vector database onto a prompt and wants to know what should go in it, what should never, how to measure whether it helps, and how to forget on purpose. Assumes Book 2 or the equivalent; basic Python; no framework, no GPU.The AI Agent Architecture Series is the architect's track: one architectural layer per book, on one running system the reader refactors and grows by hand. This is Book 3, Memory & Knowledge - what the agent remembers, designed and measured.Draw the memory map - working, episodic, profile, semantic, procedural - from the run states Atlas actually produces, and measure the day the prompt overflowsTreat the context window as a budgeted cache with priorities, relevance, an order that survives to the prompt, and an eviction log - and draw the budget curve to its kneeHand-build a vector store - tokens, vectors, cosine, three chunkers - and show with a table that how a document is cut matters more than how it is searchedRebuild RAG as a pipeline with a metric at every stage and a grounding contract that refuses an invented sentence by name - and learn exactly what the contract cannot catchAdd a knowledge graph over the systems of record for the questions that are paths, not passages - and see the failure of graphs built from proseGate every write-back by source: episodes, stated preferences, and human verdicts kept; tool results and model prose refused - then remove the gate and watch the index grow sixteenfoldReproduce a lost update a hundred times and fix it with compare-and-set: one integer, one comparison, one loop, zero lost This item is printed on demand. Shipping may be from our UK warehouse or from our Australian or US warehouses, depending on stock availability.…