Discussion and conclusions

4. Discussion

4.1 Summary of main findings

For RQ1, the evidence converges on a parallel hierarchical model: the periphery already performs non-linear computation (antagonism, gating by sniff rhythm); the olfactory bulb separates two temporal channels and is already plastic with learning; the piriform synthesises odour objects from experience and acts as an associative memory with predictive capacity; the amygdala encodes valence at the single-cell level (decoding identity from ~400 ms) and emotional salience (valence × intensity) at the fMRI level; the OFC represents subjective value, dependent on labels, state and the individual; the entorhinal–hippocampus records odour–context associations, with priority for the first association, and is modulated by dopamine. For RQ2, the characteristic behavioural effects are real but small-to-moderate (d ≈ 0.3–0.5), and considerably smaller after bias correction. For RQ3, applied claims range in certainty from moderate (TMR, Alzheimer’s marker, Silexan, retail recall), to low (olfactory training, inhaled lavender, citrus/lemon), to very low (aromatherapy for dementia, depression, sleep).

4.2 Re-evaluating the “limbic shortcut” hypothesis

The most popular hypothesis holds that odours evoke strong emotions and memories because they go straight to the amygdala and hippocampus without passing through the thalamus. The data support the anatomical premise [5, 22, 23] but not its simple corollary, for four reasons:

1. Speed does not differ as expected. The amygdala decodes odour identity only 50 ms after the piriform, but the hippocampus needs ~1.1 s [11]. Anatomically “close” does not mean informationally “fast” (a direct comparison with vision using the same method is not yet available in the dataset).
2. Odour-evoked memories are not more effective. Odours evoke memories less successfully than words/pictures (63% vs 76%) [32]. A true shortcut would predict the opposite.
3. Valence exists before the amygdala. Olfactory bulb signals recorded through the skin already correlate with valence from ~250 ms [65], and in mice innate valence has its own hard-wired circuit via the CoA [66]. Valence is therefore not an exclusive product of the amygdala (although timing comparisons between EBG and cell recordings call for caution; identity decoding in the PPC/amygdala begins from ~110 ms [51]).
4. Expectations and labels reproduce many physiological and perceptual effects without any odour [98, 106, 107] — effects that the limbic hypothesis predicts should depend on the chemical. A notable exception: lemon lifts mood regardless of expectation [105], so odour still has its own effect on some outcomes.

A more accurate description, supported by current data: odour perception is itself a memory process [15, 27, 47]. The piriform and hippocampus share an auto-associative architecture; odour objects exist only after they have been learnt; and the valence of most odours is learnt valence (the piriform–BLA–OFC pathway), overlaid on a small amount of innate valence (the CoA pathway). Odours do not “trigger” emotional memories in the sense of a key opening a door; odours are recognised through the very emotional–contextual associations that formed them. This simultaneously explains the high specificity (when evoked, the memories are deep [84]), the individuality (the OFC encodes according to the person’s own evaluation [12]; cultural differences for methyl salicylate [33]), and the poor effectiveness as a general cue (a matching association is needed).

4.3 Comparing theoretical frameworks

No single framework suffices on its own. The unified framework proposed here — memory-based, valence-laden, predictively modulated olfactory perception — retains the anatomical premise of the limbic framework, the mechanism of the predictive framework, and the phenomenological observations of the valence and LOVER frameworks.

4.4 Reconciling the contradictions

• Amygdala: intensity or valence? [52] vs [60] — resolved by the intensity × valence interaction [62] and single-cell evidence [11]. Hemispheric dissociation (unpleasant odours → right [4, 30]) exists at the meta-analytic level but is not yet stable at the level of single studies.
• Are odour memories more emotional? Hackländer [32] says yes; Willander & Larsson [85] unclear. The differences are due to sample (students vs older adults) and design (within- vs between-subjects); the pooled estimate of d ≈ 0.47 should be regarded as an upper bound.
• Does lavender work? Positive meta-analysis [101] vs negative expectation-controlled RCTs [105, 106]. Reconciliation: the effect is present in high-anxiety, unblinded situations, where expectation and the caregiver’s attention are part of the intervention; a specific pharmacological–olfactory effect has not been demonstrated.
• Single-cell recording in humans: Dikeçligil & Gottfried (2024) write “not yet available” [1] whereas Kehl 2024 had already been published [11] — a discrepancy of timing, not a contradiction in the data.
• Number of OR genes (391–~1,000): due to annotation definitions; see 3.2.

4.5 Strengths and limitations

Strengths. (1) Cross-level scope from atomic structure to consumer behaviour; (2) 80% of papers were read in full text or partial full text (63% complete), with detailed notes and structured extraction (277 effects); (3) certainty assessed per outcome rather than per paper; (4) a transparent data-error log (Appendix C) — 20+ inconsistencies in the source literature were detected and handled.

Limitations. (1) Not a comprehensive systematic search: the paper set is a pre-existing list plus targeted searches, so there may be selection bias towards highly cited work and human studies; (2) no registered protocol; (3) extraction and GRADE assessment were done by a language-model-assisted process, without two independent reviewers — mitigated by cross-checking between notes and extraction, but not a substitute; (4) 24 papers were available as abstract only, including some key papers (Gottfried 2006 [48], Anderson 2003 [52], Rasch 2007 [80]) — their figures were taken from secondary sources; (5) no new pooling, so Figure 8 is a descriptive comparison, not a meta-analysis; (6) applied data come mainly from Western populations (91% in [117]); of the 119 included papers, none comes from Vietnam or Southeast Asia.

4.6 Implications for the fragrance industry

Three neurally grounded design principles: (1) congruence of scent–context–brand matters more than absolute pleasantness [117]; (2) simplicity — multidimensional scents are weaker [117], consistent with the limit of ~3 identifiable components [46]; (3) deliberate repetition in positive moments — suggested by mouse data and by a survey showing that 73% of earliest odour memories come from repeated events [34], together with the hippocampal imprint of the first odour association [87]. Low-dose unpleasant notes may enrich the value response in the OFC (indole in jasmine [53] — no evidence yet of increased pleasantness ratings); valence asymmetry suggests that avoiding the risk of unpleasant odours matters more than maximising pleasantness [119].

4.7 Proposed research programme

Sample-size calculations (two-sided α = 0.05, 80% power, comparison of two independent groups) based on the estimates above: detecting d = 0.36 (chemosignal) requires ~121 people/group; d = 0.64 (lavender, low-risk RCTs) ~39/group; d = 0.13 (corrected TMR) ~930/group — far exceeding any existing TMR experiment (mean n 22). Four priority studies, feasible in Vietnam, are proposed:

1. Valence–familiarity norms for Vietnamese people. A multicentre survey of 40–60 odours (including local odours: fish sauce (nước mắm), agarwood (trầm), lemongrass (sả), honey locust (bồ kết), pandan leaf (lá dứa)) with GEOS [67] and a standard valence scale; n ≈ 400 per age–region subgroup to estimate means with an error of ±0.1 SD (95%). Rationale: valence is culture-dependent [33].
2. A 2 × 3 RCT: expectation × odour. Instruction (relaxation vs neutral) × odour (lavender, a pleasantness-matched control odour, no odour), with outcomes STAI-S, HRV and salivary cortisol; n ≈ 40/cell (240) is sufficient to detect d ≈ 0.64 for the main effect. This design extends [105, 106] with a pleasantness-matched control odour and clinical outcomes, in a Vietnamese population.
3. Olfactory training after viral infection with everyday Vietnamese odours vs the standard kit, randomised, with blinded assessors, over 24 weeks; TDI outcome separated into threshold vs identification to test the central-learning hypothesis [111].
4. Field experiment on brand memory: scent on the product vs ambient scent vs no scent, measuring recall at 24 hours and 2 weeks [118], with mediation analysis of the number of scent reminders.

Open fundamental questions: (a) causal manipulation of the piriform–olfactory bulb predictive loop in humans (TMS/tDCS combined with iEEG); (b) the active molecule of the fear chemosignal; (c) whether the hippocampal imprint of the “first association” [87] explains the childhood peak of odour memories; (d) adequately powered, pre-registered TMR trials.

5. Conclusions

The odour–emotion–memory link has a solid neural basis and is understood in ever greater detail, from the geometry of the receptor binding pocket to the valence code of individual amygdala neurons. However, the mechanism is not a simple “shortcut”: odour perception is a memory-based process, carries valence from very early on, and is strongly governed by expectation, labels and state. Quantitatively, the characteristic effects are small to moderate; in terms of application, moderate certainty belongs to TMR, the Alzheimer’s marker and retail recall, while memory tied to scented objects, mood-lifting by lemon and olfactory training have promising evidence but low certainty, and the therapeutic claims of aromatherapy remain largely unproven. The fragrance industry will benefit most by building products and communication on the portion of evidence with moderate or higher certainty, and by investing in studies with expectation-controlled designs in its own market.

Scientific reference material, not medical advice. Figures are redrawn from published data, not copied from the original figures. GRADE and risk-of-bias ratings are the compilers’ own. Every number traces back to detailed reading notes on each paper.