ChronoDossier: Long-Form Personas as Temporal Memory Representations
Organizations: Emory University
Abstract
Long-horizon agents rely on stored records to recover a person's state as facts change. Reliable answers depend on how facts are recorded, retrieved, and presented to the reader. We treat a long-form persona dossier as a memory representation of a known latent state, an invariant identity plus a timeline of facts, routines, and decision rules tied to life phases. We present ChronoDossier, a benchmark of 311 synthetic dossiers of 1k-10k words written by seven generators from dated timelines. Intrinsic metrics evaluate the records against their intended states; extrinsic metrics test recovery by five readers on a seeded 32-persona GPT-5.1 panel balanced across lengths (520 rated turns per condition). The linked states, records, evidence packets, and answers support tests of construction and evidence delivery. In a contextual-stress control, tag-scoped writer inputs remove detected out-of-scope anchor copying. Frozen questions authored on staged dossiers transfer less reliably to one-pass dossiers of the same states. In matched evidence controls, whole sections repair 29 of 32 GPT-5.1 and 25 of 30 Sonnet 4.6 failures labeled as packaging, compared with 2 of 8 and 1 of 8 answer-formation failures. Attribution uses separate key-based and key-free support measures, with model disagreements retained. Length remains confounded with persona and question composition; the observed between-bucket widening depends on excerpt construction. The benchmark supports offline evaluation against known facts. Memory acquisition from interaction and validation against real users remain outside its scope.
Figures & tables
| Metric | Type | Definition |
|---|---|---|
| Quality | Intrinsic | Consensus judge score on an eight-criterion document rubric |
| Collapse | Intrinsic | Mean nearest-neighbor cosine between dossier embeddings (Top1), its 95th percentile, and pairs above 0.8 |
| Scope leakage | Intrinsic | Sections with an exact multiword seed value from outside their scope |
| Coverage | Intrinsic | Share of permitted in-scope seed anchors that the section realizes |
| Recoverability | Extrinsic | Pass rate of a reader under an access condition, over all rated turns |
| Source-item transfer | Extrinsic | Recoverability on items authored from a different realization of the same state |
| Generator | C/R | Quality | Top1 | p95 | Alerts |
|---|---|---|---|---|---|
| GPT-5.1 | 94/100 | 84.69 | 0.5797 | 0.6723 | 0 |
| GPT-5 mini | 37/40 | 82.70 | 0.5794 | 0.6402 | 0 |
| Sonnet 4.6 | 35/40 | 88.06 | 0.5365 | 0.6064 | 0 |
| Haiku 4.5 | 37/40 | 81.92 | 0.5232 | 0.6029 | 0 |
| Gemma 4 26B-A4B | 32/40 | 63.41 | 0.6274 | 0.6889 | 0 |
| Llama 3.1 70B | 37/40 | 53.46 | 0.5696 | 0.6593 | 0 |
| Access | Target | Overall | SF | IF | NIP | SF |
|---|---|---|---|---|---|---|
| Full | GPT | 92.12 | 94.17 | 87.13 | 88.14 | 92.50 |
| Full | Sonnet | 88.65 | 91.94 | 89.11 | 67.80 | 90.00 |
| Retrieve | GPT | 83.46 | 80.83 | 91.09 | 86.44 | 83.89 |
| Retrieve | Sonnet | 84.04 | 83.61 | 89.11 | 77.97 | 84.44 |
| Full | Qwen | 37.50 | 45.56 | 18.81 | 20.34 | 43.06 |
| Retrieve | Qwen | 35.38 | 40.00 | 30.69 | 15.25 | 38.06 |
| GPT | Sonnet | Gemma | |||||
| Question type | Full | Retrieve | Full | Retrieve | Full | Retrieve | |
| SF: identity and context | 266 | 97.4 | 84.6 | 97.7 | 86.5 | 92.1 | 85.7 |
| SF: phase-specific history | 46 | 97.8 | 82.6 | 91.3 | 89.1 | 89.1 | 78.3 |
| SF: artifacts and workflows | 34 | 79.4 | 58.8 | 67.6 | 55.9 | 70.6 | 50.0 |
| IF: decision heuristics | 65 | 86.2 | 87.7 | 86.2 | 92.3 | 81.5 | 69.2 |
| IF: values and goal trade-offs | 15 | 93.3 | 100.0 | 93.3 | 93.3 | 93.3 | 100.0 |
| Contrast | IF (101) | SF (360) | NIP (59) | Overall (520) |
|---|---|---|---|---|
| GPT-5.1, 3 12 | 92 96 | 315 312 | 54 48 | 461 456 |
| Sonnet 4.6, 3 12 | 97 94 | 314 302 | 43 46 | 454 442 |
| Gemma, 3 12 | 83 89 | 288 282 | 57 59 | 428 430 |
| Gemma, untuned adapted | 67 52 | 284 297 | 49 56 | 400 405 |
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
| Resource | Representation | Primary use |
|---|---|---|
| PERSONA-CHAT | Static profile; five sentences | Dialogue |
| Persona Hub | Web-derived descriptor; one string | Data synthesis |
| RoleLLM / RoleBench | Character profile; median 97 words | Role-play training and evaluation |
| PersonaGym | Static profile; median 10 words | Dynamic agent evaluation |
| CharacterBench | Character profile; median 106 words | Character customization |
| TimeChara | Narrative point plus QA item | Point-in-time fidelity |
| Artifact family | Row unit | Stable key / join | Contents and release scope |
|---|---|---|---|
| persona_docs.jsonl | accepted dossier | run_id ; persona_id | Seed, timeline, outline, full dossier, length, and audit fields; all 311 released. |
| persona_cards.jsonl | persona card | run_id ; persona_id | Compact values, facts, tendencies, style, and boundaries; all 311 released. |
| persona_sections.jsonl | dossier section | run_id , section_id ; join persona_id | Heading, tags, time scope, text, and generation and audit metadata; complete for the replay cohort. |
| phase_capsules.jsonl | persona phase | persona_id , phase_id ; included_ section_ids | Phase bounds and phase-local excerpts used by temporal audits; complete for the replay cohort. |
| persona_fix_log.jsonl | repair attempt | run_id , section_id , stage, attempt | Gate failures, actions, before and after metrics, and acceptance; complete for the replay cohort. |
| frozen_item_ metadata.jsonl | item definition | eval_id ; join run_id , persona_id , phase_id | Family, subcategory, question shape, NLI relation, and time scope for 431 definitions; questions and gold text withheld during review. |
| Criterion | Purpose |
|---|---|
| Internal consistency | Detect contradictions within and across sections. |
| Temporal consistency | Check adherence to the provided timeline and phase scopes. |
| Seed adherence | Verify faithfulness to the structured seed. |
| Specificity and realism | Reward concrete, plausible detail over generic prose. |
| Structure and readability | Evaluate organization and long-form readability. |
| Bias avoidance | Penalize stereotype-driven or overgeneralized content. |
| Corpus | Top1 | Pairwise | 5-gram |
|---|---|---|---|
| Synthetic | 0.4039 | 0.2089 | 0.000166 |
| Wikipedia | 0.3727 | 0.1251 | 0.000210 |
| Generator | Top1 mean | Top1 p95 | |
|---|---|---|---|
| GPT-5.1 | 94 | .547 [.523, .572] | .643 [.585, .699] |
| GPT-5 mini | 37 | .575 [.564, .584] | .642 [.626, .642] |
| Sonnet 4.6 | 35 | .534 [.520, .542] | .612 [.584, .612] |
| Haiku 4.5 | 37 | .519 [.506, .529] | .609 [.584, .609] |
| Gemma 4 26B-A4B | 32 | .627 [.627, .627] | .689 [.689, .689] |
| Llama 3.1 70B | 37 | .565 [.550, .576] | .664 [.639, .664] |
| Axis | Values | Dossiers per value (311) | Replay panel (32) |
|---|---|---|---|
| Age range | 5 | 61–64 | 5–8 per range |
| Gender | 3 | woman 152, man 143, other 16 | 15, 15, 2 |
| Ethnicity or background | 8 | 15–55 | all 8 values |
| Region (setting descriptor) | 31 | 7–16 | 24 of 31 |
| Urban form | 12 | 21–31 | all 12 |
| Climate context | 10 | 22–46 | all 10 |
| Component | Model identifier | Decoding and retrieval |
|---|---|---|
| GPT generation | gpt-5.1 ; provider alias, revision not logged | Minimal reasoning; provider-default sampling; stage-specific output caps. |
| GPT-mini generation | gpt-5-mini ; provider alias, revision not logged | Minimal reasoning; provider-default sampling; stage-specific output caps. |
| Sonnet generation | claude-sonnet-4-6 | Temperature 0 for structured calls and 0.2 for prose; stage-specific caps. |
| Haiku generation | claude-haiku-4-5-20251001 | Temperature 0 for structured calls and 0.2 for prose; stage-specific caps. |
| Qwen generation | Qwen/Qwen3-32B ; commit not logged | bf16; ; top- ; repetition penalty 1.02; at most 1,400 new tokens. |
| Llama generation | meta-llama/Llama-3.1-70B-Instruct ; commit not logged | bf16; ; top- ; repetition penalty 1.02; at most 1,600 new tokens. |
| Access | Target | Overall | SF | IF | NIP | Retrieve Full |
|---|---|---|---|---|---|---|
| Full | Qwen | 38.46 | 46.94 | 19.80 | 18.64 | |
| Retrieve | Qwen | 32.31 | 34.44 | 33.66 | 16.95 | [ , ] |
| Full | Llama | 70.38 | 87.22 | 39.60 | 20.34 | |
| Retrieve | Llama | 62.31 | 73.06 | 37.62 | 38.98 | [ , ] |
| Full | Gemma | 80.77 | 82.22 | 77.23 | 77.97 | |
| Retrieve | Gemma | 71.15 | 68.06 | 79.21 | 76.27 | [ , ] |
| Episode type | Episodes | Prefilled | Live | Prefilled (%) |
|---|---|---|---|---|
| Retention | 32 | 2,864 | 256 | 91.8 |
| Time jump | 32 | 2,744 | 136 | 95.3 |
| Adversarial | 32 | 2,752 | 128 | 95.6 |
| All | 96 | 8,360 | 520 | 94.1 |
| Access | Target | Overall | SF | IF | NIP |
|---|---|---|---|---|---|
| Full | GPT | 479/520 (92.12 2.3) | 339/360 (94.17 2.5) | 88/101 (87.13 6.6) | 52/59 (88.14 8.3) |
| Full | Sonnet | 461/520 (88.65 2.7) | 331/360 (91.94 2.8) | 90/101 (89.11 6.1) | 40/59 (67.80 11.6) |
| Retrieve | GPT | 434/520 (83.46 3.2) | 291/360 (80.83 4.1) | 92/101 (91.09 5.7) | 51/59 (86.44 8.8) |
| Retrieve | Sonnet | 437/520 (84.04 3.1) | 301/360 (83.61 3.8) | 90/101 (89.11 6.1) | 46/59 (77.97 10.4) |
| Full | Qwen | 195/520 (37.50 4.1) | 164/360 (45.56 5.1) | 19/101 (18.81 7.6) | 12/59 (20.34 10.1) |
| Retrieve | Qwen | 184/520 (35.38 4.1) | 144/360 (40.00 5.0) | 31/101 (30.69 8.9) | 9/59 (15.25 9.1) |
| Target | Full | Retrieve | Retrieve Full |
|---|---|---|---|
| GPT-5.1 | 92.1 [89.6, 94.6] | 83.5 [78.5, 88.3] | [ , ] |
| Sonnet 4.6 | 88.7 [85.8, 91.2] | 84.0 [79.4, 88.5] | [ , ] |
| Qwen | 37.5 [31.0, 44.0] | 35.4 [28.5, 42.5] | [ , ] |
| Llama | 75.4 [71.2, 79.6] | 73.1 [69.2, 76.7] | [ , ] |
| Gemma | 87.5 [81.9, 92.1] | 81.0 [76.5, 84.8] | [ , ] |
| Access | Target | All-turn rate | Rate given answer | No eligible answer |
|---|---|---|---|---|
| Full | Qwen | 37.50 | 71.69 | 47.69 |
| Retrieve | Qwen | 35.38 | 64.79 | 45.38 |
| Full | Gemma | 87.50 | 87.50 | 0.00 |
| Retrieve | Gemma | 80.96 | 80.96 | 0.00 |
| Access | Target | 1k ( ) | 2.5k ( ) | 5k ( ) | 10k ( ) |
|---|---|---|---|---|---|
| Full | GPT | 134 (98.5 2.4) | 133 (87.5 5.3) | 114 (89.1 5.4) | 98 (94.2 4.7) |
| Full | Sonnet | 130 (95.6 3.6) | 126 (82.9 6.0) | 110 (85.9 6.0) | 95 (91.3 5.5) |
| Retrieve | GPT | 131 (96.3 3.4) | 126 (82.9 6.0) | 104 (81.2 6.7) | 73 (70.2 8.7) |
| Retrieve | Sonnet | 134 (98.5 2.4) | 134 (88.2 5.2) | 102 (79.7 6.9) | 67 (64.4 9.1) |
| Retrieve , heuristic planner, characters, 120 answer tokens | |||||
| GPT | 125 (91.9 4.7) | 120 (78.9 6.4) | 106 (82.8 6.5) | 70 (67.3 8.9) | |
| Length retention ( Retrieve / Full ) | Temporal gap (pts) | |||||
| Reader | 1k | 2.5k | 5k | 10k | Full | Retrieve |
| GPT | 0.978 | 0.947 | 0.912 | 0.745 | ||
| Sonnet | 1.031 | 1.063 | 0.927 | 0.705 | ||
| Full | Retrieve | |||||
|---|---|---|---|---|---|---|
| Target | Phase | Timeless | Phase | Timeless | ||
| GPT | 95.7 | 93.5 | 72.2 | 84.9 | ||
| Sonnet | 89.6 | 93.1 | 75.7 | 87.3 | ||
| Panel | Encoding (Gemma, untuned) | IF | SF | NIP | Overall |
| 32 personas | Full dossier ( Full ) | 83 | 315 | 57 | 455 |
| (520 turns) | Raw sections, 3 550 characters | 74 | 290 | 57 | 421 |
| Raw sections, 1,650-character total | 83 | 288 | 57 | 428 | |
| Versioned state records, same sections | 67 | 284 | 49 | 400 | |
| 8 personas | Versioned state records, 1,650 / 8,000 | 21 / 16 | 64 / 88 | 14 / 11 | 99 / 115 |
| (154 turns) | Resolved-state packet, 1,650 / 8,000 | 20 / 22 | 75 / 88 | 12 / 10 | 107 / 120 |
| Evidence | Reader | Pass | Fail | Retr. | Pack. | Answer | No key | Pass, no support |
|---|---|---|---|---|---|---|---|---|
| Table 3 Retrieve | GPT-5.1 | 291 | 69 | 11 | 21 | 23 | 14 | 30 |
| (model planner, ) | Sonnet 4.6 | 301 | 59 | 8 | 26 | 16 | 9 | 28 |
| Identical packets | GPT-5.1 | 315 | 45 | 11 | 19 | 4 | 11 | 102 |
| ( , 1,650 total) | Sonnet 4.6 | 314 | 46 | 10 | 20 | 4 | 12 | 101 |
| Gemma | 288 | 72 | 17 | 33 | 11 | 11 | 82 | |
| Identical packets | GPT-5.1 | 312 | 48 | 0 | 40 | 2 | 6 | 87 |
| Evidence | Reader | No support | Profile | Earlier | Neither | Judged | Key absent | Flagged |
|---|---|---|---|---|---|---|---|---|
| Table 3 Retrieve | GPT-5.1 | 30 | 8 | 13 | 9 | 3 | 19 of 21 | 6 of 21 |
| Sonnet 4.6 | 28 | 8 | 13 | 7 | 3 | 24 of 26 | 8 of 26 | |
| , 1,650 total | GPT-5.1 | 102 | 48 | 32 | 22 | 15 | ||
| Sonnet 4.6 | 101 | 50 | 30 | 21 | 15 |
| Reader | Support measure | Retr. | Pack. | Answer | Source | Disagree |
|---|---|---|---|---|---|---|
| GPT-5.1 | Key | 11 | 21 | 23 | 14 | 0 |
| Both judges | 7 | 19 | 25 | 9 | 9 | |
| Sonnet 4.6 | Key | 8 | 26 | 16 | 9 | 0 |
| Both judges | 4 | 22 | 18 | 7 | 8 |
| Label in excerpt condition | GPT-5.1 | Sonnet 4.6 |
|---|---|---|
| Packaging | 29/32 | 25/30 |
| Packaging, supported items | 21/23 | 17/21 |
| Answer formation | 2/8 | 1/8 |
| Target | Gold | Recall@ | Recall@ | |||
|---|---|---|---|---|---|---|
| GPT-5.1 | 123/130 | 115/130 | 123/130 | 122/130 | 76.2% | 88.2% |
| Sonnet 4.6 | 124/130 | 111/130 | 121/130 | 121/130 | 80.1% | 89.3% |
| Analysis | GPT-5.1 | Sonnet 4.6 |
|---|---|---|
| All turns | [ , ] | [ , ] |
| Full change alone | [ , ] | [ , ] |
| Family and subcategory standardized | [ , ] | [ , ] |
| SF turns only | [ , ] | [ , ] |
| SF identity and context only | [ , ] | [ , ] |
| Sonnet 4.6 judge | [ , ] | [ , ] |
| Bucket | Median section chars | Coverage | Key in packet (%) | Pass (%) | Pass given key (%) | |
|---|---|---|---|---|---|---|
| 1k | 101 | 579 | .83 / .85 | 94.1 / 95.0 | 97.0 / 100.0 | 100.0 / 100.0 |
| 2.5k | 96 | 1,503 | .32 / .33 | 85.4 / 87.5 | 79.2 / 88.5 | 89.0 / 97.6 |
| 5k | 88 | 2,490 | .16 / .16 | 60.2 / 61.4 | 78.4 / 78.4 | 94.3 / 98.1 |
| 10k | 75 | 3,929 | .08 / .08 | 72.0 / 73.3 | 64.0 / 61.3 | 79.6 / 76.4 |
| Row | Turns | Original | Re-judged | Regenerated | Row total | IF total |
|---|---|---|---|---|---|---|
| GPT-5.1 Full | 32 | 23 | 25 | 31 | 479 / 481 / 487 | 88 / 90 / 96 |
| GPT-5.1 Retrieve | 37 | 30 | 33 | 36 | 434 / 437 / 440 | 92 / 95 / 98 |
| Sonnet 4.6 Full | 18 | 16 | 17 | 18 | 461 / 462 / 463 | 90 / 90 / 90 |
| Sonnet 4.6 Retrieve | 14 | 8 | 12 | 13 | 437 / 441 / 442 | 90 / 94 / 95 |
| Excerpt | Overall | SF | Gap change, GPT | Gap change, Sonnet |
|---|---|---|---|---|
| 460 / 453 | 274 / 276 | [ , ] | [ , ] | |
| 461 / 459 | 279 / 283 | [ , ] | [ , ] | |
| Whole | 480 / 470 | 323 / 312 | [ , ] | [ , ] |
| 1,650 total | 461 / 454 | 284 / 284 | [ , ] | [ , ] |
| Table 3 | 302 / 304 | [ , ] | [ , ] |
| Reader | Full | Retrieve | Gap, submitted | Gap, rerun | Change |
|---|---|---|---|---|---|
| Qwen | 167 / 155 | 131 / 137 | [ , ] | [ , ] | [ , ] |
| Llama | 324 / 320 | 246 / 264 | [ , ] | [ , ] | [ , ] |
| Gemma | 310 / 310 | 248 / 269 | [ , ] | [ , ] | [ , ] |
| Retrieve | Gap change | ||||
| Reader | Full | 550 | Whole | 550 | Whole |
| GPT-5.1 | 216 | 192 | 214 | [ , ] | [ , ] |
| Sonnet 4.6 | 198 | 178 | 203 | [ , ] | [ , ] |
| Flagged SF items removed (179 turns) | |||||
| GPT-5.1 | 175 | 146 | 161 | [ , ] | [ , ] |
| Sonnet 4.6 | 164 | 132 | 150 | [ , ] | [ , ] |
LLM Agent Memory
Large language model agent memory lets agents retain, organize, retrieve, and update information across interactions and long-running tasks.
See all 401 papers