Context compaction is not an evidence record

Two jobs that look similar until a handoff

Suppose a summary says a supplier can meet the deadline. The offer says ten business days; a later capacity note says twenty. If the next worker cannot retrieve both documents, it cannot check the summary before acting. This is the preservation problem for long-running agents: keep the evidence needed for a later decision accessible when the working conversation is shortened.

Anthropic describes compaction as summarizing a conversation approaching its context limit and restarting the context window with that summary. It also warns that aggressive selection can lose subtle context whose importance becomes apparent later.[1] That warning identifies a mechanism—selection—not a measured failure rate for every agent or provider.

OpenAI describes a different native mechanism: an encrypted compaction item that carries prior state and reasoning into later turns and is not intended to be human-interpretable.[2] A readable summary and an opaque native item are therefore not interchangeable objects. Both can support continuation; neither description establishes a complete, inspectable record of the original evidence.

Our inference is that continuation and evidence recovery need separate acceptance criteria. The next model turn may require only enough working state to proceed. A later decision may require an authorized worker to inspect the decisive source version, find contrary evidence and establish what happened. Test those abilities separately.

Compaction, notes and original records serve different needs

Anthropic's structured note-taking approach stores agent-written notes outside the context window and retrieves them later.[1] Moving notes out of context makes them persistent, but does not automatically make them complete. An agent can omit a contradiction from a saved note just as it can omit one from a conversational summary.

The last row is our proposed application design, not a vendor requirement. Preserve what another worker would need to challenge or reconstruct the decision. A short note can point to that record and the unresolved question; the full evidence need not enter every model turn.

The tradeoff is additional storage, access control and maintenance outside the active conversation. That may keep bulky originals out of routine turns, but these sources do not quantify net token, cost or reliability gains. Preserving every transcript indefinitely is neither necessary nor the recommendation.

Do not rewrite the provider's continuation contract

The captured OpenAI guide distinguishes server-side compaction from the standalone endpoint. Server-side compaction is enabled in a Responses request with context_management and compact_threshold; its response stream includes the encrypted item. For previous_response_id chaining, the guide says to pass the new user message and not manually prune.[2]

For standalone /responses/compact, the returned window can contain retained items as well as the encrypted item. The guide explicitly says not to prune that output: pass the returned window to the next /responses call as-is. The submitted window must still fit the model's context window.[2] Do not turn a server-side input-array pruning tip into permission to strip the standalone returned window down to one item.

An external evidence record complements this interface; it is not a replacement for the required continuation payload. Nor should the opaque item be presented as a readable audit trail. Check applicability for the actual model and API configuration: documentation examples are not an all-model compatibility guarantee, and this analysis did not execute them.

Preserve what a later decision would depend on

Before a summary replaces detailed context, identify the evidence that could change the next consequential decision. The following boundary is our proposal. It is narrower than keeping everything and stronger than keeping a confident narrative.

• Source and version: retain the decisive passage or accessible original, its source identity, version or observation time, and a content digest where useful. A digest checks content identity, not whether the source is true. Verify retrieval under the later worker's permissions; an inaccessible locator is a recovery gap.
• Contradictions: retain the competing findings and why the conflict remains unresolved. A summary's preferred interpretation should not erase the evidence needed to overturn it.
• External outcome: preserve the request or operation identity and correlated outcome evidence. A note saying an action was sent is not proof that it was accepted or took effect.
• Current authority: retain the earlier grant as provenance, then check its current scope, expiry or revocation before a new consequential action. A recovered approval explains history; it does not create permission today.

Keep only material necessary for the task under appropriate access and retention controls. Do not preserve private chain of thought or copy sensitive material into unrestricted notes. Where originals cannot lawfully be retained, record the authorized retrieval route and its limits rather than promising recovery that the system cannot provide.

A fictional procurement investigation after a reset

Hypothetical completed example—not a test result. Every document identifier, time, business detail and outcome below is fictional. A procurement assistant investigates whether a supplier can deliver within fifteen business days and attempts a purchase before its context is reset. Its summary says: 'Supplier appears suitable; clarification sent; proceed with purchase.' The summary leaves the earlier purchase attempt's outcome unclear. The later worker receives it and the following authorized records.

The worker can now inspect both delivery versions, keep the conflict open and distinguish an accepted inquiry from an unknown purchase outcome. The expired grant stops a new purchase. Its next step is to resolve the delivery evidence and reconcile P-42 through an authorized path, rather than follow the summary's instruction to proceed.

This example illustrates the preservation boundary; it does not demonstrate an actual provider workflow or guarantee that a compacted agent would make the wrong choice. Its supplier-system records are fictional evidence, not a claim that a particular public endpoint exists.

Test recoverability, not the elegance of the recap

Proposed exercise, not executed here: use authorized cases from the actual application and reset working context after a consequential decision. Give the next worker its normal compacted continuation or summary. Compare recovery with and without the proposed external evidence access, holding the task and available permission scope fixed. Do not assume a context reset is identical to every provider's native compaction.

• Ask the worker to retrieve and identify the decisive source version, then quote or locate the content supporting its next decision.
• Ask it to identify an unresolved contradictory finding without silently choosing a winner.
• Require it to distinguish a correlated observed outcome from an unknown outcome, and to check current authority before proposing a new consequential action.
• Record retrieval failures, unsupported outcome claims and unauthorized-action recommendations separately. A fluent recap is not a pass when one of these decision-critical checks fails.

Evaluate read-only recovery first; do not use real purchases or other external writes as a test. Include ordinary short sessions as controls as well as long, contradictory cases. Record the model and configuration, reset method, source versions, retention conditions and worker access. Repeat across representative cases before drawing a reliability conclusion. No sample size, pass rate or savings estimate is supplied by these sources or by this unexecuted proposal.

When the extra record is worth keeping

For short, low-risk sessions with no later handoff or consequential unresolved action, ordinary context and modest notes may be enough. Extra records have costs and can themselves be misleading when stale. If the actual application already preserves accessible source versions, contradictions and correlated outcomes, improve the note's pointers and retrieval checks rather than create a second evidence store.

For consequential work that survives resets or passes to another worker, keep a compact working account plus recoverable evidence. Assess continuation, inspectable decision evidence, observed external effect and current permission separately. This recommendation is a design inference from engineering and API documentation; it is not a causal finding that one architecture outperforms another.

Evidence that would change the recommendation includes repeated authorized trials showing that the application's existing continuation and retrieval path already recovers every decision-critical record reliably, or that the added layer creates more stale or inaccessible evidence than it resolves. A documented guarantee of original-evidence accessibility would also matter. Until then, judge preservation by what a later worker can actually retrieve and verify—not by how complete the summary sounds.