Memory lifecycle

How Engram maps operational AI memory events to new memory, active context, and stable knowledge.

Engram uses a brain metaphor to make memory operations legible, but the underlying data model is operational rather than anatomical.

Operation	System meaning	Engram visualization
store	A durable record was created	New memory in the hippocampus
retrieve	The memory system returned or selected candidates	Retrieval path toward working memory
load	The application placed memory IDs into model input	Active context in the prefrontal cortex
update / supersede	A fact changed or an older record became stale	Versioned memory transition
summarize	Multiple records became a semantic summary	Stable knowledge in the temporal cortex

Retrieval is not context

A memory provider can return ten candidates while the prompt builder loads only two. Engram therefore requires separate evidence:

await turn.retrieve({
  query,
  candidates,
  selectedIds: ["memory-current-city"]
});

await turn.load(["memory-current-city"]);

retrieve proves which records the application received or selected. load proves which records the instrumented application says it placed into model context.

Corrections preserve history

When a user says, “Actually, I live in Oakland now,” a robust memory system should not silently erase all evidence that San Francisco was previously stored. It should preserve the update relationship and prefer the current fact during retrieval.

Engram represents that distinction with update and supersede operations. Incident branches can test a corrected state without mutating the original provider.

Consolidation creates stable knowledge

Repeated or related episodic records may be summarized into a semantic memory. Engram records the source and target IDs:

await turn.summarize(
  {
    id: "memory-bay-area-history",
    content: "User has lived in the Bay Area and currently lives in Oakland."
  },
  ["memory-san-francisco", "memory-oakland"],
  "Consolidated location history"
);

The visualization maps this semantic target to stable knowledge. It does not claim that production AI memory systems literally reproduce human temporal-lobe behavior.

The anatomical labels are an educational interface. Memory Telemetry v2 contains tiers, scopes, operations, and evidence metadata—not brain regions.