Most intelligent work leaves a poor record of how it happened.
A notebook contains conclusions whose failed paths have vanished. A codebase retains the patch but rarely the uncertainty that produced it. An AI conversation generates proposals, objections, calculations, and small discoveries, then compresses them into a final answer or loses them when the context ends.
Artifacts sometimes survive. Their paths usually disappear.
An organization can preserve every document, message, model checkpoint, and execution trace while remaining unable to answer a simple question: how did we arrive here, and where could we resume differently?
Useful memory must expose the paths among those records.
Paths between artifacts
Suppose an evaluation reveals a particular failure in a candidate. A modification produces a child. That child imports a mechanism from another branch, encounters a different environment, and is rejected because it violates a risk constraint.
A folder can contain all of those artifacts. A lineage records the operations between them: evaluation, counterexample, mutation, composition, another evaluation, and the reason for rejection.
The edges matter as much as the nodes. They distinguish mutation from correction, composition from coincidence, and failure under one environment from failure everywhere.
One can move backward from an outcome to the evidence that authorized it, sideways into alternatives available at the time, or forward into descendants that attempted to resolve the same counterexample. A dormant branch can return when new constraints make it relevant.
Replay
Most records preserve one view. A chart preserves one projection of some data. A video preserves one camera path through a scene. A report preserves the conclusions selected by its author.
A replayable artifact retains the transitions beneath the view: source evidence, state updates, available behaviors, authorization, action, and consequence.
The renderer then becomes a choice. The same path might appear as a chart, an animation, a simulation, a research report, or the context supplied to another agent. A human may watch a compressed episode while an agent inspects every event relevant to one decision.
This is why the language of keyframes is useful outside animation. A keyframe marks a consequential state or transition. Another observer can reconstruct the path at a different resolution without changing what occurred.
Conditions travel with evidence
An evaluation belongs to a candidate under stated conditions: environment, time window, observation mode, execution model, costs, and constraints. Its evidence may include several objectives, behavioral descriptors, uncertainty, provenance, and the path that produced the outcome.
A final score is one reduction of that record.
This connects lineage to the shape of better. A Pareto frontier depends on the objectives and constraints used to construct it. Preserving the underlying evidence lets a later lens ask a different question without rewriting the earlier evaluation.
A trading policy rejected under taker fees may become useful on a venue with zero maker fees. A model rejected for latency may become viable on different hardware. A scientific method below one instrument’s noise floor may work when measurement precision changes. The evidence stayed fixed while the decision context moved.
What to keep
Preserving structure still permits aggressive deletion and compression.
An archive that retains every token, intermediate table, duplicate measurement, and rendered projection becomes harder to use as it grows. The durable record can be much smaller. It needs identity, parentage, the operation that produced it, evaluation conditions, primitive outcomes, disposition, source references, and enough information to replay or recompute what matters.
High-volume traces can be compressed. Derived views can be regenerated. Invalid evidence can retain a compact failure witness.
Several distinctions must survive compression. Missing evidence is different from numeric zero. Superseded evidence retains the conditions under which it applied. Rejection records the lens that produced it. A summary can reduce confidence in its source; it cannot silently increase it. Memory and eligibility to act remain separate.
Counterfactuals
Ordinary organizational memory mostly records what was selected. Selection removes information: the winning proposal hides its alternatives, the deployed model hides candidates that disagreed, and a final action hides the policies that abstained.
A structured population preserves some of those alternatives. It can recover what else was available, why it was not selected, what it would have done under the same evidence, and which later observation changed that judgment.
Some unrealized branches require new simulation or measurement. Recording where that uncertainty lives makes memory operational: it can propose the next experiment.
Continuation
AI agents make the loss of lineage unusually visible. They produce useful intermediate work quickly, but much of it remains trapped inside conversations, temporary files, and local context. The next agent often begins by rediscovering the problem.
A traversable lineage changes the handoff. The agent receives the current object, its active alternatives, the evidence that changed them, unresolved counterexamples, and the operations available next.
In a map made of reactions, a repertoire of behaviors becomes an instrument for perceiving the world. Lineage gives that instrument a history. Replay lets another observer enter it. Lenses let the same evidence answer a new question.
At the current rate of candidate generation, this continuity supplies the heredity required for a population to evolve. That is the subject of after mutation became cheap.