Educational companion dossier · Fact, interpretation, lived experience, clinical education, fiction, and mechanics are labeled separately. Scope & safety
PsychologicalWar.orgPeople before labels.
REAL-WORLD INTERPRETIVE

AI-DRIVEN WORLD SAFETY

Memory Bootstrap, Retrieval, Supersession, and Forgetting

A bounded memory model with stable IDs, event and learned dates, provenance, confidence, privacy, revision, contradiction, retrieval, and correction.

Read the working brief Choose a learning path
LEVEL 1

ORIENTATION

Why this matters

REAL-WORLD INTERPRETIVE

One-sentence brief

Unlimited transcript replay produces cost, context rot, privacy exposure, and identity drift. A selective memory system should preserve continuity without pretending that every line is equally important or permanently available.

REAL-WORLD INTERPRETIVE

Three key points

  1. Bootstrap memories are structured records, not a hidden monologue.
  2. Retrieval balances relevance, recency, importance, privacy, and current contradiction state.
  3. Corrections supersede old memories rather than silently rewriting history.
LEVEL 2

WORKING BRIEF

Evidence, context, and limits

REAL-WORLD INTERPRETIVE

Author a memory bootstrap

Create a small set of autobiographical, relationship, commitment, preference, world-fact, rumor, emotional, scene, correction, and future-intention memories with stable IDs and provenance.

  • Seed ordinary life, not only plot clues.
  • Do not create memories of player actions that never occurred.
REAL-WORLD INTERPRETIVE

Separate event date from learned date

A character may learn about an earlier event later. Store both the time of the event and the time and source of acquisition to support chronology and knowledge access.

  • Hearsay and witnessing differ.
  • Confidence may change after correction.
REAL-WORLD INTERPRETIVE

Retrieve a bounded working set

Rank eligible memories by relevance to the current input and scene, recency, importance, relationship, privacy, and contradiction state. Apply diversity and recency constraints so one theme does not dominate every turn.

  • A retrieval score is not truth.
  • Private or forbidden memories remain excluded.
REAL-WORLD INTERPRETIVE

Correct and supersede

When a memory is disproven or refined, create a correction or new revision and point the old record to its supersessor. Runtime uses the current record while audit retains the earlier belief.

  • Belief can change without erasing history.
  • Propagate correction to derived summaries.
REAL-WORLD INTERPRETIVE

Model forgetting and persistence carefully

Some scene details can decay; commitments, core relationships, and high-impact events may persist. Retrieval policy should not simulate disability, trauma, or cognitive impairment through stereotypes.

  • Forgetting is a design parameter, not a diagnosis.
  • Players need predictable correction and deletion rules.
LEVEL 3

COMPLETE DOSSIER

Limitations, game links, and review context

DISPUTED / MULTIPLE ACCOUNTS

Known limitations and gaps

  • The six supplied reports are preserved research leads. Their filenames, organizations, citations, examples, thresholds, and technical detail do not authenticate authorship, sponsorship, product status, or current factual accuracy.
  • The public section is design literacy, not a production specification. It does not expose provider credentials, prompts, live endpoints, activation tokens, autonomous tool use, or implementation code for a running agent system.
  • Population-quality measures must detect mechanical repetition and coherence failures without treating demographic rarity, disability, nationality, language, religion, gender, migration, or another protected characteristic as a defect or quality score.
  • Thresholds, similarity methods, sampling plans, language rules, and runtime budgets require representative testing, privacy review, accessibility review, cultural and linguistic review, security review, and accountable human approval before any production use.
REAL-WORLD INTERPRETIVE

Decision matrix

Memory eligibility.
Memory state Retrieve? Reason
Current, relevant, permitted Yes, within budget Supports continuity
Private from current participant No Knowledge and disclosure boundary
Superseded or disproven No; retrieve correction instead Avoid stale belief as current fact
Low-confidence rumor Maybe, labeled Character may discuss uncertainty
Administrative review note Never Not character memory
REAL-WORLD INTERPRETIVE

Publication audit checklist

Identity and agency

Does the design preserve the exact fictional identity, ordinary life, independent goals, and ability to refuse rather than reducing the character to a role or prompt?

Pass condition: Identity fields are stable, state is separate, protected traits are not quality scores, and silent substitution is impossible.

Evidence and review

Can every transition, validation result, accepted fingerprint, exception, and human decision be traced to a versioned record?

Pass condition: Automated checks, human review, activation authority, and production approval remain separate and explicit.

Runtime boundary

Can untrusted provider output, administrative evidence, stale revisions, or private data enter live context or binding state?

Pass condition: Only allowlisted, current, reviewed projections and bounded scene or memory packets can be used; failures degrade safely.

Correction and retirement

Can a changed source, identity revision, harmful behavior, or failed review invalidate downstream use without destroying audit history?

Pass condition: Supersession, pause, rollback, correction, and permanent retirement are defined and testable.

LEVEL 4

RESEARCH EDITION

Sources, methods, and stable links

REAL-WORLD VERIFIED

Method and corrections

This page follows the public method for provenance, confidence, source independence, alternative accounts, limitations, review state, and visible correction.

Read the methodView corrections
NEXT

CONTINUE

Related learning

AI-Driven World SafetyAI Mission Generation: Reliability, Verification, and Fail-ForwardA mission pipeline in which generative systems propose bounded narrative variations while deterministic validators ensure that objectives, locations, rewards, permissions, and exits are real.AI-Driven World SafetyAI NPC Cost, Latency, and Graceful DegradationA bounded architecture for generative characters that protects budget, responsiveness, privacy, and authoritative game state under normal play, spikes, outages, and abuse.AI-Driven World SafetyAI NPC Identity, Consent, Memory, and ProvenanceA clear boundary for AI-driven characters: players know when generative systems are involved, what data is used, what the NPC remembers, and which outputs can change game state.AI-Driven World SafetyBatch Correlation, Ordering, Idempotency, and RetryA deterministic transport contract that maps every character response to one request, preserves order, handles retries safely, and keeps identity separate from network metadata. Learning pathRuntime Projection, Memory, and ReliabilityKeep creator evidence, accepted projections, scene state, memory, prompt context, budgets, fallbacks, identity, and incident response separate.