The source stayed the same. The agent's job changed.

An agent has a well-prepared packet: source excerpts, a summary, useful distinctions, and a recommendation. Reusing it avoids repeating expensive reading and reasoning.

Then the mission changes.

The recipient is different. An earlier constraint is withdrawn. The work moves from exploration into commitment. A new instruction changes which parts may be shared.

The source files can remain byte-for-byte identical while the packet needs another interpretation.

Keep the layers visible.

Source material says what was available to inspect.

An interpretation says what an earlier task made relevant.

A prepared contribution packages that interpretation for a purpose and recipient.

These layers can be reused differently. The source may remain useful even when a recommendation no longer fits. An explanation may survive while its proposed next action changes. A private working note may inform new writing without being suitable for public delivery.

Treating the entire packet as one reusable object can hide those distinctions.

Token savings are one part of the account.

Reusing prepared context may reduce repeated acquisition, summarization, or model processing. It may also require retrieval, validation, adaptation, and correction.

What matters is the cost of reaching an acceptable mission result. A shorter context can still omit a condition that the task depends on. A longer one can carry irrelevant or outdated instructions that increase the work of interpretation.

The comparison belongs in total reuse cost, with a baseline aimed at the same finish.

A practical review before depending on a packet

Ask what the packet was prepared to do; which sources and conditions it consumed; which interpretation remains useful; and what changed in the receiving task.

That review does not require redoing every earlier step. It can identify the part that needs reconsideration while retaining unaffected source work.

A hash match can establish matching bytes. It cannot establish that two missions mean the same thing. Similarity likewise supplies a clue, not a complete permission or fitness decision.

Reuse the right contribution.

An agent may retain the source summary, revise the recommendation, and compose a new public explanation. It may reuse a prepared calculation while checking a changed input. It may choose fresh work because reconstructing confidence in the packet would cost more.

Those are different reuse decisions. Each deserves an account of what it preserves and what it reopens.

Read selective renewal, equivalence, and the definition of a finish. For development-team applications, explore agent builders or team learning.

Good memory makes earlier work available. Good judgment determines what that work is still good for.