Graph Engineering - The Loop Memory Model

Give agentic systems a memory that survives every run by turning decisions, evidence, and execution history into durable, traceable graph structuresKey FeaturesSeparate work history from durable knowledge with a two-graph architecture and clear bridge rulesRetrieve small, provenance tracked context windows from graph paths within token budgetsRun parallel workers with scoped promises, validated writes, reducer contracts, and hard budgetsBook DescriptionAgent loops are good at producing work, but the reasoning behind that work disappears the moment the context window closes. Two agents rediscover the same bug three sprints apart; a decision made in run 12 gets silently contradicted in run 47.Graph Engineering moves that reasoning out of transcripts and into an explicit knowledge graph. You'll model execution history separately from durable knowledge, then connect the two with a small set of bridge relationships. Markdown stays the authoritative source; the graph is a disposable index built for queries, not a second source of truth.From there, the book turns the model into a working mechanism: parse structured repository data, extract only genuine prose, resolve entities, and assemble typed relations. Retrieve small, relevant neighborhoods instead of dumping an entire knowledge base into context. See how evaluations check supporting paths, how concurrent workers publish changes without corrupting shared state, and how a pipeline preserves lineage from objective to artifact.The final section is judgment; which metrics reveal graph quality, how false merges and bad incentives produce confident errors, what this architecture costs, and when a simple ADR log or relational table is the better answer.What you will learnChoose the right architecture from loop, chain, swarm, DAG, or graphDefine separate schemas for work history and in-force knowledgeTurn Markdown nodes into a disposable query indexAdd provenance, confidence, and validity to graph assertionsCompute NEAR links to surface duplicates without making them factsBuild a context assembler with typed, time-aware traversalCheck claims against graph paths instead of prose aloneMeasure graph quality and know when not to build oneWho this book is forThis book is for senior software engineers, AI and agent engineers, software architects, and platform teams already experimenting with agent loops, coding agents, or multi-agent workflows. It is most useful when decisions, constraints, evidence, and execution history need to survive across runs or be shared between workers. Prior books in the Clarity series are helpful but not required; the manuscript recaps the lattice before extending it. Readers should be comfortable reasoning about software architecture and structured engineering workflows.