An agent works out something hard (a migration order, a flaky-test cause, a non-obvious constraint), and then the session closes and it is gone. Virgo commits those learnings to the graph, so the next agent and the next engineer start from them instead of rediscovering them.
Behind one question, the flake’s whole world lights up: the test, the fixture, the CI config, and the two earlier reports nobody linked. The conclusion writes back to the graph.
Every agent commits what it learns to the same graph, with provenance, so a finding from one session is already context for the next, in any tool.
Committed once, a learning serves the next engineer, the next agent, and the next quarter.
The flake, its cause, and its fix stay bound to the test and the config that produced them. The third occurrence costs a lookup, not an afternoon.
An agent that forgets everything at the end of each session can only ever be as good as its last prompt. One that writes learnings to the graph, with provenance, makes every solved problem context for the next one.
Roughly 3x fewer iterations in code generation when the relevant memory is on hand instead of rediscovered.
Built on Agent memory → and Agents grounded in real context →