When context wears out

Long AI projects rarely fail in one dramatic moment. They become unclear slowly. An early assumption remains in the conversation after it has been rejected. A summary replaces the source from which it was derived. New requirements are added without retiring old boundaries. Eventually the system knows many sentences about the project, but no longer knows its current state reliably. Context has not disappeared. It has worn out.

Context is not volume; it is selection

In AI work, abundant context is often treated as a synonym for good preparation. That is only half true. More material initially increases the chance that important information is present. It also increases the number of competing instructions, outdated decisions and seemingly relevant details. A system cannot infer what is binding today when everything carries the same weight in its working space.

Good context therefore answers a question of selection rather than quantity: Which information changes the next decision? A complete chronology may be valuable for documentation and still be unsuitable for the task at hand. Working context must be smaller than the archive that feeds it.

This distinction creates two spaces. The archive preserves history, sources and earlier variants. Active context contains only the objective, current decisions, authoritative materials, unresolved tensions and the next verifiable step. When the spaces are mixed, every new conversation must separate past from present again.

Three properties determine context quality

The first is relevance. Information belongs in active context when it changes the desired outcome, a boundary, an acceptance criterion or the next action. Interesting background is not enough. Relevance is task-specific and may change from one phase to another.

The second is provenance. It must remain visible whether a statement comes from an authoritative file, a human decision, a model summary or an untested assumption. Without provenance, claims become uniform. An elegant summary can quietly receive more authority than the material it compressed.

The third is compactness. Compact does not mean shallow. It means that every included item has a function. Repetition, historical explanation and superseded variants are not destroyed; they move to the archive. Active context remains readable enough for people and systems to notice contradictions.

Saturation begins before the technical limit

Context can wear out long before a technical window is full. The first warning is often not an obvious error but a change in response quality: requirements are followed only partially, language becomes generic, resolved questions return or new proposals conflict with earlier decisions.

Saturation is therefore also a problem of attention. The more rules are simultaneously visible, the harder they are to prioritise. A model can repeat an old instruction correctly and still apply it in the wrong place. People experience something similar: an extensive project description feels complete even when nobody can name the three points that matter now.

Do not wait for collapse. Define early maintenance signals: repeated corrections, lost exceptions, conflicting terminology, unclear sources or the feeling that each new task requires the whole project to be explained again. When two or three signals appear together, the context needs controlled distillation, not another addition.

Silent drift revives old decisions

Context wear often grows through well-intentioned continuation. A summary is updated, summarised again and later used as the basis for another hand-off. Conditions, uncertainty and reasoning can disappear at every stage. The text becomes shorter and apparently clearer while its connection to the original decision weakens.

Superseded decisions are particularly dangerous. An old variant may return later when its invalidity was not recorded as visibly as its introduction. A project therefore needs more than a list of what applies. It needs a small decision history: replaced, rejected, valid until or still open.

Drift can also be linguistic. “Should be evaluated” becomes “is planned” in a later summary and finally “will be implemented”. The content seems stable while its level of commitment changes. Status words are not a stylistic detail. They prevent possibility, intention and decision from collapsing into one another.

Phases need a deliberate ending

A long chat should not continue merely because it feels familiar. Each project phase asks different questions and needs different material. Research gathers possibilities, design chooses, implementation needs binding specifications and review needs criteria plus observable results. When every phase remains in the same uncleaned context, their modes of thinking compete.

A phase ending has three movements. First, decide: What applies, what was rejected and what remains open? Second, distil: Which few materials carry the next phase? Third, hand over: What outcome should be created next, how will it be tested and when must work stop?