AI contextual memory

Definition

Contextual memory is what lets a writing assistant not start from scratch on every interaction. It stores and organizes project-specific elements: character sheets, magic or technology rules, chronology, target tone, summaries of scenes already written, research notes. Without this layer, an LLM only "remembers" the current conversation window — enough for a one-off rephrase, insufficient for a 100,000-word novel where chapter 32 must honor a promise from chapter 3.

Technically, it often combines several mechanisms: session memory (short term), RAG (retrieval over a document base), and persistent structures such as the Narrative knowledge graph. Narrative entity extraction feeds this memory by turning raw text into named entities and usable relationships. Memory is not a full copy of the manuscript: it is a selective representation, kept current and queryable by the Narrative copilot (AI).

What it's for

This is the #1 criterion for telling a general-purpose tool from an author-focused one. With reliable memory, the assistant can answer "does Léa already know her mother's secret?" based on actual scenes, not guesswork. It reduces involuntary contradictions between spaced writing sessions and enables Internal coherence analysis across the whole project.

For the author, contextual memory mainly offloads mental load: less need to remember everything, more confidence to move forward without rereading 400 pages before writing the next chapter. It does not replace careful rereading — it complements the factual and structural part of it.

Variants by genre

• Fantasy / SF: rules + ontology memory — magic systems, species, technology, maps; contradictions often involve abilities or distances.
• Mystery: chronological and epistemic memory — who knows what, when; alibis, planted clues, red herrings; facts must be timestamped.
• Romance: relational memory — emotional states, implicit promises, history of confrontations; sensitive to characterization regressions.
• Literary fiction: light on plot, rich on voice — register, recurring motifs, references; less factual graph, more stylistic profile.
• Multi-volume series: cross-book memory with versioning — separate what is canon from book 1 from what was retconned in book 3.

Common pitfalls

• Assuming a long prompt equals memory: pasting 50 pages into a chat does not create persistent, queryable memory.
• Memory mixing multiple projects: leakage between manuscripts, characters, and world rules.
• Memory without user control: no way to forget a false lead, fix a mis-extracted entity, or archive an old draft.
• Stale memory: base not updated after major rewrites — the assistant cites obsolete versions of the story.
• Over-trusting RAG: retrieving out-of-context passages that mislead instead of clarifying.
• No Narrative entity extraction: raw textual memory, hard to query for relationships ("who is allied with whom?").

Examples

• Dune (world bible) — Herbert maintained notes on politics, economics, and genealogy; modern contextual memory would structure these elements to feed a copilot without writing in his place.
• A Song of Ice and Fire — character and alliance density makes relational memory essential; a graph-based system would map lineages and oaths.
• The Murder of Roger Ackroyd — epistemic memory (who knows what, including the narrator) is crucial; unstructured AI memory would miss the twist by blending knowledge levels.

Writing tips

• Check what the tool actually "sees": a 128k-context LLM doesn't use 128k usefully — often the first 20 and last 20 pages; the middle is fuzzy.
• Feed hierarchical summaries (chapter → book → saga) rather than raw text: a good 500-word synthesis beats 30,000 words dumped in.
• Purge memory between two projects: fantasy concepts can pollute a thriller if the AI keeps parasitic context.
• Stay judge of the memory: never delegate continuity verification to AI memory alone. Cross-check with your written codex.

Learn more: Complete guide: novel coherence on 100,000+ words and Narrative threads: stop forgetting your story's promises.

Revision checklist

• Can I show AI contextual memory in action in at least two scenes without stating it in a "theme" sentence?
• Do Narrative knowledge graph and this notion reinforce each other, or does one contradict the other?
• Could a beta reader name AI contextual memory after reading without being told?
• Am I keeping a scene out of attachment that doesn't illustrate this notion?
• Did my last pass fix structure before style (see Line editing / polishing if needed)?