Two views with different responsibilities
Saved history records the exchange that occurred. The working context supplies what the model needs to continue within its limits. Their separate jobs allow reduction of the working view without replacing the learner’s words with generated summaries in the transcript.
This is not merely a storage detail. If a summary becomes the history, the original statement and its conversational position disappear. A generated interpretation of an error could become the only evidence for it. Reduction should remain a derived reading whose original can be revisited.
Why a summarisation setting is not enough
Reduction operates alongside persistence, tool calls and session restoration. A component suitable for service-owned history may be unsuitable where the application manages its own history. A feature name in a library does not demonstrate that it works in our composition.
Warda’s decision record documents a re-examination after discovering an integration mismatch. We retained the intended outcome—continuing dialogue without replacing history—and changed where reduction is applied. Adopting a library means checking its composition, not simply invoking an API with the expected name.
A tool call and result belong together
An agent may request a lesson and receive its content. Preserving the request while removing its result leaves an incomplete exchange. Preserving an unexplained result creates a different problem. History management must respect relationships between messages instead of deleting a fixed number of old lines.
We also distinguish bulky repeated material from recent details about a learner’s question. Space consumption does not determine educational value. A useful reduction retains a coherent basis for the next turn while keeping the original exchange available.
A useful summary does not acquire authority
A summary can be wrong or carry traces of retrieved content. We do not promote it to system-instruction authority or treat it as independent proof of mastery. Summarisation is itself an AI operation with source and interpretation boundaries, even when it runs outside the visible conversation.
Exam state has a different policy. Questions, answers and their order contribute to the integrity of judgement; they are not compressed into a free-form summary. A tutoring policy does not automatically transfer to examinations simply because both use sessions.
A summary has a known position in history
A summary cannot freely replace an arbitrary part of history. Its covered range must be known, followed by newer messages without deleting an attempt absent from the summary or reinserting old material in a confusing position. These boundaries can be checked against a defined transcript.
Background work introduces another case: a second request may arrive or the conversation may advance before reduction finishes. A summary’s relationship to its covered range matters more than arrival time alone. A later timestamp does not necessarily mean wider coverage. We check composition before preparing the view; saved history remains the reference for every derived reading.
What context review examines
Context review examines repeated material, continuity, the arrival of a new summary and the continued presence of a bound lesson. It considers extended dialogue and the read path, rather than merely admiring a shorter text.
The tradeoff requires checks that preserve original messages and prevent synthetic context from entering saved history. These contracts make memory management part of platform engineering. They do not establish perfect memory or demonstrate improved student learning on their own.
Sources and context
- Warda engineering decision — read-path context reduction
Reduction decision and September context-reconstruction study. Session records and measurements are not published.
Internal records document Warda’s decisions and engineering contracts. External references, where included, explain a pattern or specification; design quality and effects on student learning are different questions.