Designing Long-Form Roleplay Memory

Keep long stories coherent by recording what changes future decisions and discarding transcript noise.

By HereHaven Team · 6 min read · Published · Updated

Example files

Read the current Guide Book

Start a Lore-Guided Chat with a small canon file

Lore is retrieved as relevant entries; the complete file is not necessarily present in every prompt. Private creator Lorebooks are not a public reference library. These downloadable examples contain no private user records.

  • Download one Harbor Lorebook format below. The supported extensions are .txt, .md, and .json, with a maximum file size of 250 KB (256,000 bytes). A PDF, archive, or screenshot is not a supported Lorebook upload.
  • Start a new character chat and choose Lore-Guided Chat. Under Pick the Lorebook for this chat, use Upload your own Lorebook override to select the file. A creator Lorebook may already be selected; an upload replaces that source for this setup.
  • Wait for extraction to finish. The success state says Lorebook parsed successfully. If setup fails, read the error before using Try again. Remove blank content, check the extension and size, and validate JSON syntax rather than repeatedly sending the same broken file.
  • Ask a natural recall question: What must we bring to open the tide archive? The sample canon says a brass token is required. Then ask about the blue chart and check whether the response preserves its copied status.
  • Record correct facts, omissions, and invented details in the test worksheet. File validation only verifies that the upload is readable; it does not establish extraction or model accuracy. Resolve conflicting canon before expecting reliable recall.

Separate memory horizons

Long-form roleplay needs durable memory, scene memory, and disposable conversational detail.

Assign each new fact a horizon when it appears so permanent identity does not compete with temporary positions or phrasing.

When everything is permanent, important information becomes difficult to retrieve and contradictions multiply.

Use three separate note blocks during a HereHaven story. Canon holds facts expected to survive chapters; scene state holds positions, objects, and immediate offers; scratch space holds theories and wording that can expire. At the end of a turn, place each new detail in exactly one block. This prevents a temporary rainstorm from competing with an oath, while keeping the current lantern position available until the characters leave the courtyard.

Record durable facts

Durable facts should influence later choices or establish canon that would be costly to contradict.

Store names, commitments, revealed rules, lasting consequences, ownership, and resolved mysteries in short declarative lines.

Do not preserve speculation as truth; label uncertain claims and identify their source.

Write durable facts as subject, state, source, and confidence. “North gate opens at dawn; observed by user and Orin; confirmed” is safer than “gate opens.” A disputed note can say “Orin claims keeper owns duplicate key; unverified.” When later evidence resolves uncertainty, replace or supersede the note rather than stacking both versions. Source labels let the story preserve lies, mistakes, and partial knowledge without converting them into narrator truth.

Track current scene state

Scene state answers where everyone is, what they can observe, and what action is underway right now.

Update location, participants, relevant objects, immediate objective, obstacles, and open offers after each meaningful turn.

Stale scene state causes teleporting objects, duplicated actions, and responses to questions already settled.

Update scene state only after an action completes. If Orin reaches for the map, record an open attempt, not possession; the user still controls whether the transfer occurs. Track the current holder of important objects, participant locations, immediate hazards, unanswered questions, and elapsed deadline pressure. This small ledger prevents accidental consent, duplicate actions, and spatial jumps while giving the next HereHaven reply concrete material to continue.

Measure relationship change

Relationships become convincing when trust, obligation, rivalry, or affection changes because of specific events.

Record the cause and direction of a change rather than assigning an unexplained score.

A single friendly exchange should not erase betrayal or create lifelong loyalty.

Describe relationship movement with cause, direction, and remaining limit. “Trust rises because the sealed key was returned; Orin still withholds family history” preserves both progress and restraint. Avoid numeric scores unless they correspond to explained behavior. Before changing a relationship note, identify the scene event that earned it and one future choice it affects. This keeps warmth, rivalry, obligation, and suspicion connected to play rather than arbitrary mood changes.

Use recap discipline

A recap is a control surface for the next scene, not a literary summary of everything that happened.

Write five compact lines: achieved, changed, learned, unresolved, and next pressure; verify each against the story.

Avoid adding motives or facts during recap that no scene established.

Write recaps from state changes rather than transcript order. Scan the completed scene for objectives achieved, facts confirmed, possessions transferred, relationship shifts, and unresolved pressures. Each recap line should answer what a future reply must know. Exclude abandoned plans and atmospheric details unless they caused a lasting effect. For the archive scene, footsteps matter only while present; the copied-key mystery and Orin’s debt survive because both can redirect decisions in the next chapter. Read the recap alone and verify it supports a coherent, playable next-scene opening without inventing a missing transition or unsupported outcome. Ask which single note would prevent the most damaging continuity error, and make that note easiest to find.

Resolve memory conflicts

Conflicting notes require source checking before either version becomes canon.

Prefer direct scene outcomes over character guesses, preserve intentional uncertainty, and correct obsolete state explicitly.

Silently keeping both versions makes future responses alternate between incompatible realities.

Review at scene boundaries

Boundary reviews prevent accumulated notes from becoming an unfiltered second transcript.

Promote lasting consequences, carry forward unresolved threads, archive completed scene positions, and remove redundant wording.

The final memory should help write the next reply without requiring the full chat.