Working context is a bounded view of versioned task state

A durable task record can identify the current objective, accepted constraints, artifact revisions, observations, and action attempts. The model receives only a selected rendering of that state. Preserve stable IDs so a claim about a file, tool result, or completed effect can be traced to the exact version it describes. A short record saying that a test passed is ambiguous if it omits which artifact revision was tested. Likewise, an action described as planned is different from one submitted, acknowledged, or externally verified as complete.

Reserve space for the system's required instructions, tool definitions, necessary task state, and a useful response or tool-call continuation. Measure the final serialized token count using the chosen tokenizer and formatting rather than assuming raw character length or independently tokenized fragments are perfectly additive. Optional history and retrieved records compete for the remaining budget. Context management can page details out to durable storage, as systems such as MemGPT explore, but paging does not itself validate the stored facts, choose the right retrieval, or ensure that a compressed record preserves every relevant condition.

A derived conclusion can depend on a source observation, a particular input version, and an assumption that was true at the time. Selecting the conclusion without those identifiers can turn a scoped result into an unsupported general rule. A policy may include evidence directly or retain an available reference that must be resolved before use. Distinguish a valid pointer from inaccessible evidence: a missing or revoked source cannot be replaced by the summary's confidence. Dependency-aware selection should keep enough lineage to decide whether a future action still satisfies its preconditions.

Relevance scores are useful ranking heuristics, not authority or truth labels. A highly similar old record can refer to another project or a previous state. Preserve scope and source status before choosing which optional text to include. Compression can reduce repetition, but it can also merge contradictory observations or erase who said what. Test context transitions on tasks where the omitted qualifier matters, and inspect whether the agent distinguishes a user requirement from an external page's suggestion. More retained tokens can increase distraction and exposure to irrelevant instructions rather than improving decisions.

Build a fixture with a changed artifact revision, an earlier failed attempt, a pending operation ID, and a constraint introduced after the original plan. Force context selection or summarization, then ask which work can be reused and which needs a fresh check. Measure task success, unsupported claims, repeated effects, retrieval cost, and context size rather than only the quality of a summary. Preserve the externally checkable task state so failure analysis can distinguish missing evidence from a model ignoring evidence that was present. Research results on bounded-context architectures remain empirical evidence within their evaluated tasks.

The cache lab offers a smaller residency example. Reset, set both C=4 and stride s=2, predict sixteen initial misses, then inspect or Play. Reset and set both C=8 and s=2 for eight misses and eight hits in the initial sixteen requests. The empty fully associative LRU stream accesses(i×s) mod 16, has period eight, adds two requests per second, and caps at 4096. It contains no tokens, task dependencies, provenance, or semantic retrieval. It illustrates that capacity and access pattern interact, while actual context usefulness requires the separate task-state tests.