The canonical result is the published paper:
arXiv:2608.28609. This file maps its main
claims to the committed outputs and distinguishes them from earlier exploratory
trained-memory experiments that remain in results/ for research provenance.
- Text is a lossy perceptual-memory channel. A caption-based re-identifier retains as little as 0.11 of a dedicated encoder's recall on non-nameable signals.
- Grounding and identification solve different problems. On the two-person face task, whole-scene encoding reaches 0.05 recall; VLM grounding followed by alignment and ArcFace reaches the correct-region oracle at 0.96.
- The in-model read faithfully reproduces the encoder. With the final sharp, training-free read, AttMem and cosine retrieval make the same decision on random and hard-distractor banks.
- The result is portable. The encoder match holds across ten frozen model families and a five-model-by-five-modality grid.
- Perceptual and text memory compose. Perceptual identity recalls a native marker; that marker can key exact facts in a text store.
Both methods below use identical registrations and queries over 20 draws, with the target marker slot randomized on each draw.
| Regime | Cell | Encoder cosine | AttMem | Difference |
|---|---|---|---|---|
| random | Face, N=10 |
0.948 | 0.948 ± 0.033 | 0.0 pp |
| random | Face, N=100 |
0.792 | 0.792 ± 0.021 | 0.0 pp |
| random | Face, N=300 |
0.749 | 0.749 ± 0.018 | 0.0 pp |
| random | Face, N=1000 |
0.776 | 0.777 ± 0.008 | 0.0 pp |
| random | Style, N=5 |
0.473 | 0.473 ± 0.104 | 0.0 pp |
| random | Style, N=10 |
0.428 | 0.428 ± 0.076 | 0.0 pp |
| hard distractors | Face, K=19 |
0.853 | 0.853 ± 0.007 | 0.0 pp |
This equality is intentional. The contribution is the grounded, in-model container and its composition behavior; recognition quality belongs to the chosen encoder.
- 12-domain breadth: 1,080 PerceptMem tasks at
N=10,20,40. The in-model memory tracks each domain's encoder and exceeds chance in every domain. - Model universality: every cell in the five-model-by-five-modality grid is within one recall point of its encoder; most are within 0.003.
- Open set: verification AUROC is 0.99 for voice and 0.97–0.99 for faces. Stranger rejection is 0.96–0.99 at 20 enrolled identities and 0.81 at the harder 80-user agent operating point.
- Composition: face-to-fact accuracy is 4–10× the face-withheld chance baseline and follows recognition accuracy times the model's in-context lookup reliability.
- End-to-end agent: at 80 enrolled users, identity and routed fact recall are both 0.86, stranger rejection is 0.81, and overall task success is 0.83.
- Capacity router: compressing
Midentities intokprototype slots followsmin(1, k/M) × C(M), whereC(M)is the encoder ceiling. Exact facts instead fail from key/value binding interference and belong in a text store.
| Claim or figure | Primary committed outputs |
|---|---|
| paired encoder match | attmem_paired_*.json |
| grounding faces | agentic_prod_*.json, agentic_realistic_*.json |
| grounding paintings | agentic_paint_*.json |
| grounding audio | agentic_audio_*.json, conversation_grounding_*.json |
| text-caption baselines | text_baseline.json, text_baseline_audio.json, text_baseline_style.json |
| cross-domain benchmark | cross_domain.json |
| in-model composition | composition*.json, reasoning_eval.json |
| open-set evaluation | openset_verification.json |
| end-to-end agent | agent_benchmark.json |
| latency and cost | latency*.json, cost.json |
| modality isolation | attmem_mixed_modal.json, av_fusion.json |
Files whose names include nonzero training steps or adversarial-training probabilities record an earlier learned, softer read. Some of those experiments appeared to beat raw cosine on selected cells, but the final paired protocol identified marker-slot bias and under-sharpening as confounds. They are retained to make the research path auditable; they are not the published headline.
When a table here and a historical JSON file differ, use the paper's paired protocol and final training-free outputs.