Conversational Context: Control How AI Remembers, Adapts and Evolves

Conversational Context is the information from previous turns that continues to influence how a generative system interprets new requests within the same session. In this handout, we will examine conversation history as a source of continuity, allowing definitions, objectives, tone, constraints and decisions to persist across multiple exchanges. We will then analyze contextual inertia, where earlier assumptions, styles or goals continue shaping the interaction even after the task has changed.

Finally, we will explore context compaction, which reduces a long conversation to a selective and updated representation, and context migration, which transfers that cleaned context into a new session while preserving what remains relevant and excluding obsolete or noisy information.

Conversational Context

Fullscreen Download

Key points

Conversational context = information from previous dialogue turns made available to a model, influencing how subsequent requests are interpreted and answered.

Conversational continuity → enables reuse of established definitions, decisions, instructions, and corrections, supporting progressive work without repeatedly specifying the same information.

Contextual inertia = tendency of previously established assumptions, objectives, definitions, and stylistic conventions to influence subsequent responses, even when requirements change.

Objective inertia → previously established activities, such as brainstorming or exploration, may continue influencing generation after the task shifts toward selection or verification.

Conversational context ≠ neutral memory: accumulated history may preserve useful information alongside outdated assumptions, unresolved errors, conflicting instructions, and irrelevant details.

Visible chat history ≠ execution context: applications may supply previous messages in full, summarize them, selectively retrieve information, or exclude older content.

Context management strategies → full history inclusion, summarization, selective retrieval, message removal, and external memory integration, potentially combined within the same application.

Context presence ≠ contextual influence: information may remain available to the model without being recognized, prioritized, or correctly applied during generation.

Actual removal ≠ apparent forgetting: actual removal makes information unavailable, while apparent forgetting occurs when available information fails to influence the response.

Contextual salience → depends on factors such as clarity, repetition, explicit priority, recency, summarization, conversation structure, and application-specific history management.

Diagnostic questions → can reveal the model’s apparent understanding of previous decisions and constraints but cannot reliably expose the exact technical context supplied.

Local context rewriting = explicit reorganization of current instructions, decisions, definitions, and assumptions to eliminate obsolete information, conflicts, and ambiguity.

Context rewriting → retains valid conclusions, corrects erroneous premises, revokes outdated instructions, clarifies objectives, and distinguishes verified facts from provisional hypotheses.

Context compaction = transformation of an extensive conversation into a shorter, structured, updated representation containing information relevant to future work.

Context migration = transfer of compacted information into a new conversation, preserving project continuity without reproducing the entire previous dialogue.

Compaction ≠ migration: compaction selects and reorganizes essential information, while migration introduces that information into a new conversational environment.

Migration summary → records current objectives, valid decisions, definitions, constraints, relevant sources, completed results, unresolved questions, and information to disregard.

Complete migration ≠ targeted migration: complete migration transfers a project’s general state, while targeted migration includes only information relevant to a specific new task.

Effective compaction → preserves information according to future relevance, avoids converting hypotheses into decisions, and requires checking before migration.

Periodic compaction → conversation → updated state summary → new developments → revised summary, maintaining context as an evolving project representation.