gianlucamazza/emotional-memory-benchmarks

Emotional Memory Benchmarks

Benchmark sets for affect-conditioned memory retrieval from the emotional-memory library (Affective Field Theory, AFT). They are controlled research probes assembled by the author. They are not a production leaderboard and not a human-evaluation corpus.

• Author: Gianluca Mazza. I take LLM systems into production: state, recovery, eval, cost.
• Code and full study record: github.com/gianlucamazza/emotional-memory
• Live demo: gianlucamazza/emotional-memory-demo
• Licence: CC-BY-4.0 for every file in this dataset. The library code is MIT. Some sets were drafted with OpenAI models (noted per set below); the author releases them under the same CC-BY-4.0 licence.
• Attribution: please credit the dataset with the citation below and the concept DOI 10.5281/zenodo.19972258.

Read this first: scope and limits

• Built to favour the method being tested (Addendum U). In realistic_recall_v2, 62.5% of queries (125/200) are constructed so that affect can break a semantic tie, and the aggregate AFT advantage is confined to that regime. A win on these sets is not evidence that AFT is better in general.

• Oracle affect. The headline realistic_recall_v2 results use preset valence/arousal at encode and query time. With automatic appraisal on affect-free variants, AFT does not beat cosine (FAIL).

• Negative results recorded in the project:

  • LoCoMo (factual conversational QA): AFT underperforms a naive RAG baseline (F1 0.168 vs 0.271). FAIL.
  • MADial-Bench (third-party): cosine is significantly better (nDCG@5 0.304 vs 0.221). FAIL, inverted: the benchmark rewards supportive, counter-congruent recall.
  • ES-MemEval / EvoEmo (third-party): cosine is significantly better (nDCG@4 0.284 vs 0.133). FAIL, inverted: the gold answer is set by content, not affect.
  • Italian and Spanish with multilingual-e5 at declared power (N=120): FAIL. French (native, 2-session runner) is a PASS.

  Sources: the preregistration closures under benchmarks/ and docs/research/claim_validation_matrix.json.

• Not included here: EmoBank, MADial-Bench, ES-MemEval/EvoEmo, the DailyDialog-derived personas and LoCoMo. They are third-party or third-party-derived and keep their own licences in the GitHub repository.

• Related papers: The project also ran AFT against three published third-party benchmarks: LoCoMo (arXiv:2402.17753), MADial-Bench (arXiv:2409.15240) and ES-MemEval (arXiv:2602.01885). This dataset does not redistribute them. Their results are the FAILs listed above.

Configurations

Config	Language	Size	Origin
affect_reference_v1	en	258 examples	Script-generated from seed texts in generate_dataset.py. Whether the seed texts were drafted with an LLM is not recorded.
realistic_recall_v1	en	50 scenarios, 100 queries	Hand-authored / scripted. LLM involvement is not recorded.
realistic_recall_v2	en	50 scenarios, 200 queries	Hand-authored. Affect labels hand-crafted by the author with AFT in mind. Main benchmark.
realistic_recall_v2_it	it	30 scenarios, 120 queries	Scenarios 1–20 translated (human or LLM: not recorded). Scenarios 21–30 LLM-assisted (gpt-4.1-mini), calls logged in the repo.
realistic_recall_v2_es	es	30 scenarios, 120 queries	Scenarios 1–20 translated (tool not recorded). Scenarios 21–30 translated from the Italian top-up.
realistic_recall_v2_fr	fr	30 scenarios, 120 queries	Hand-authored in native French.
realistic_recall_v3	en	125 scenarios, 500 queries	v2 plus 75 new scenarios assembled by script. How the new texts were drafted is not recorded.
realistic_recall_v3_noAF	en	50 scenarios, 200 queries	v2 with affect labels removed.
realistic_recall_v4_noAF	en	40 scenarios, 160 queries	LLM-assisted (gpt-5-mini), author-blind.
realistic_recall_v5_gate	en	50 scenarios, 200 queries	LLM-assisted (gpt-5-mini), author-blind; 100 affective / 100 affect-free queries.
shared_history_recall_v1	en	12 scenarios, 14 sessions, 36 queries	Hand-authored, no generator script. Interleaved encode/query timelines (see below).

Full provenance per set, with commits: benchmarks/datasets/DATASET.md.

Structure

• affect_reference_v1: one row per example with id, text, valence, arousal, dominance, expected_label, source.
• realistic_recall_*: one row per scenario, with scenario_id, description and sessions[]. Each session holds events[] and queries[]. The raw JSON files in data/ also keep the top-level name, version, description and license fields that the emotional-memory runner expects.

The files are byte-identical to benchmarks/datasets/ in the GitHub repository at commit 985c0264, except shared_history_recall_v1.json, which is byte-identical at commit d6947ed5.

from datasets import load_dataset
ds = load_dataset("gianlucamazza/emotional-memory-benchmarks", "realistic_recall_v2", split="test")

shared_history_recall_v1

What it tests. Longitudinal shared-history replay. Each session declares a timeline that interleaves encodes and queries, so some memories are encoded after earlier recalls: 24 encodes-after-recall counted across whole scenarios, 22 counted within a single session. 14 of the 36 queries are near-neutral (|valence| < 0.2, share 0.3889) and 4 sit exactly at |valence| = 0.2. The set tests how affect-gated retrieval behaves when the query carries little affect.

Size and origin. 12 scenarios, 14 sessions, 60 events, 36 queries. English, hand-authored, no generator script.

Intended use. The opt-in --shared-history runner of emotional-memory, protocol mode shared_history_timeline. It is not a comparative-runner dataset, and it is distinct from the #170 shared-history replay on realistic_recall_v1.

Structure. One row per scenario (scenario_id, description, sessions[]), the same shape as the realistic_recall_* configs. Each session holds session_id, description, events[] (memory_id, content, valence, arousal, metadata{topic, phase}), queries[] (query_id, query, expected_memory_ids[], challenge_type, state{valence, arousal}) and timeline[] (kind: encode or query, id), kept in declared order. The raw JSON in data/ also keeps the top-level fields the runner expects: name, version (1.0), license (CC-BY-4.0), description, default_top_k (2), interleave (true), query_affect_gate_tau (0.2) and the declared neutral/boundary query counts.

Reference run. Produced at emotional-memory @ d6947ed5, hash embedder (TokenHashEmbedder, dim 64), seed 7, 2000 bootstrap resamples, tau 0.2, CPU, offline. Write-back is the default; --read-only-retrieves checks that retrieval leaves the store and state unchanged.

Receipt. The CI run Shared-history receipts #40084829068 reproduced these numbers at commit ce2e6579. Benchmark code and dataset are unchanged since d6947ed5. Reports: arm3_timeline_writeback.json and arm4_timeline_readonly.json, each with its .protocol.json (sha256 of shared_history_recall_v1.json: f748e37f4d7184e8979824873159738eedab661eb11c969bb7b1d42d0a617858). TokenHashEmbedder dim 64, seed 7, with both write-back and --read-only-retrieves runs. The receipt files (SHA256SUMS, receipt.json and the arm reports) are committed at benchmarks/realistic/receipts/shared_history_recall_v1/ce2e6579.

Scores: mechanism test (hash embedder), not a semantic-quality benchmark. Reference run above, n = 36, 95% percentile-bootstrap CIs.

Mode	Arm	top1 [95% CI]	hit@k [95% CI]
write-back	cosine	0.6944 [0.5278, 0.8333]	0.8056 [0.6667, 0.9167]
write-back	gated	0.7222 [0.5833, 0.8611]	0.8611 [0.75, 0.9722]
write-back	always_on	0.6944 [0.5278, 0.8333]	0.8056 [0.6667, 0.9167]
read-only	cosine	0.6944 [0.5278, 0.8333]	0.8056 [0.6667, 0.9167]
read-only	gated	0.7222 [0.5833, 0.8611]	0.8889 [0.7778, 0.9722]
read-only	always_on	0.6944 [0.5278, 0.8333]	0.8333 [0.6944, 0.9444]

Difference from cosine, with its 95% CI:

Mode	Arm	top1 diff [CI]	hit@k diff [CI]
write-back	gated	0.0278 [0.0, 0.0833]	0.0556 [0.0, 0.1389]
write-back	always_on	0.0 [-0.0833, 0.0833]	0.0 [-0.1111, 0.1111]
read-only	gated	0.0278 [0.0, 0.0833]	0.0833 [-0.0278, 0.2222]
read-only	always_on	0.0 [-0.0833, 0.0833]	0.0278 [-0.1111, 0.1667]

Result. On this set, gated and always_on show no significant difference from cosine: every CI above includes 0. Read-only left the store and state unchanged across 36 queries × 3 arms. Write-back changes top-k on 5 of 36 queries for gated and always_on, and never for cosine.

Caveats.

• The hash embedder has no semantics, so this run only tests the mechanism. It says nothing about retrieval quality.
• Query affect is a recorded oracle label; nothing infers it from the text.
• It is a small set with wide CIs.

Licence. CC-BY-4.0 for the data; the library code is MIT.

Acknowledgement. This set was motivated by an independent public review of the benchmarks by @dipankarsarkar on Hugging Face.

from datasets import load_dataset
ds = load_dataset("gianlucamazza/emotional-memory-benchmarks", "shared_history_recall_v1", split="test")

Citation

@software{mazza_emotional_memory,
  author    = {Mazza, Gianluca},
  title     = {emotional-memory: Affective Field Theory for LLM Memory},
  doi       = {10.5281/zenodo.19972258},
  url       = {https://doi.org/10.5281/zenodo.19972258},
  publisher = {Zenodo}
}