Human hippocampus represents space and time during retrieval of real-world memories

PNAS 112(35):11078–11083, doi 10.1073/pnas.1507104112; PMC4568259 (free, but the PMC, Europe PMC and PNAS copies are behind bot checks). Read for research direction R1, 4 October 2026. Provenance: papers/amadeus/nielson2015_lifelog_recall.provenance.json.

What was read

• Read in full: the 6-page publisher PDF (PNAS typeset version, PDF created 20 August 2015), from an openly hosted copy on a Max Planck graduate-school reading list. Converted with pdftotext -layout; every line was read, including Methods, figure captions and references. The two-column layout interleaves in the text file; the reading followed the columns.
• Not read: the SI Appendix (Figs. S1–S7, Table S1, SI Results: the visual-cortex comparison, and the recency, proximity and contiguity controls). It sits behind the same bot check (HTTP 403) and was not pursued.

Design

• Participants: 10 recruited from campus notice boards, aged 19–26 (mean 21.4). Nine wore the device for about a month and one for 2 weeks. One man left the fMRI session through discomfort, so 9 people were analysed, all women. Pay was $10 per day of wearing and $15 per hour of fMRI. Ohio State IRB approved; written consent was taken twice, before lifelogging and before scanning.
• Lifelogging:
  • Capture: an Android phone in a pouch on a neck strap, worn from morning to evening. A custom background app captured images, time, obfuscated audio, GPS, accelerometer and orientation, with sampling that could be triggered by movement to save battery. The average was 5,414 ± 578 images per person.
  • Privacy controls: participants could delete data, switch the app off, or cover the lens at any time.
  • Upload: nightly over SFTP while the phone charged.
• Labels from the person:
  • Episodes: on a private web page, participants segmented their own image stream into episodes and gave each a tag from a menu, a title and a short description.
  • Weekly test: after two weeks, a weekly discrimination test was run on their own images; its results are reported elsewhere.
  • Missing GPS: where GPS failed indoors, the experimenters recovered the location from the participant's description and by travelling to the place until the scene was found.
• Recall under fMRI:
  • Cues: 120 of each person's images, chosen to span the whole wearing period and to work as distinct cues. Each was shown for 8 s with the instruction "try to remember the event depicted in the picture, and try to relive your experience mentally".
  • Ratings: remembered yes/no, and vividness ("lots of detail" or "very little detail").
  • Runs: 8 runs of 15 images; 3T, 2.5 mm voxels, TR 3 s.
• Analysis:
  • Neural distance: single-trial betas; neural distance = 1 − correlation between the voxel patterns of two trials.
  • Pairs: only recalled events (63.4% ± 4.7 recalled). Pairs closer than 100 m (GPS reliability) or 15.6 h (the overnight gap) and pairs beyond 30 km were excluded, leaving 76.1 images and 1,995.7 pairs per person on average.
  • Model: per person and region, a GLM of neural distance on log time, log space, their interaction, and log time between the two presentations in the scanner (nuisance).
  • Statistics: permutation test with 10,000 permutations; Bonferroni over 10 regions (four MTL regions per hemisphere and two primary visual control regions). Researchers were not blinded.

Main results (verified)

• Left anterior hippocampus: activity patterns during recall were more dissimilar for events farther apart in real space (corrected P = 0.021), in real time (P = 0.010), and in their interaction (P = 0.038). The range was 100 m to 30 km and 15 h to 1 month.
• Not recalled, no effect: events the person did not recall showed no significant correlation.
• Specific to that region: the effect was stronger in the left anterior hippocampus than in the right anterior hippocampus or in either primary visual cortex. The authors use this against the explanation that image properties drive it.
• Not recency or proximity: neither predicted whether an event was recalled. Contiguity between consecutive cues helped recall only when same-event pairs were included.
• Data: "available via the corresponding author's institutional website".

Limits

• Small and narrow sample: nine people, all young women, from one campus.
• Group-level result: a regression slope across about 2,000 pairs per person. No single memory is decoded, and nothing about content, people or affect.
• Coarse resolution: 2.5 mm voxels and atlas regions, so no hippocampal subfields. Few within-event images, so within-event structure is untested.
• Selected cues: the experimenters chose the cues for distinctiveness, and only 63% were recalled. The neural effect conditions on the person's own report of remembering.
• Supplementary controls unread: the controls for image similarity are in the unread SI; the main text reports only their conclusions.

What it means for Amadeus

• A month of a healthy person's life can be captured passively and labelled (verified). Ordinary paid volunteers wore a camera, GPS and motion sensor daily for a month. The labels came from the person segmenting and titling their own day's stream each evening, a lifelog version of experience sampling for what happened, not for what they thought. The privacy controls (delete, pause, lens cover) and the nightly upload are a ready template. The paper does not discuss bystanders captured in the images.

• The lifelog supplies what the person cannot, the historical ground truth (inference). The P6 triangulation of the architecture needs "what they did" independently of "what they say". A lifelog gives time, place and scene for each episode, so recall under recording can be scored against the record rather than against the narration. That separates "held by the person" from "historically supported" (brief 5.3).

• Recall under fMRI reads the structure of a person's own month, not its content (verified). Distances between hippocampal patterns track real distances in space and time across a month of events. That is evidence that a person's recent autobiographical memories carry a measurable geometry, and that lifelog cues (8 s each, 120 per session) reliably drive recall. It is not a decoder, and at n = 9 it is a group effect. The architecture's hope for per-memory "variables" (salience, genuineness) is not shown here.

• Design lesson for E4 and a month-long study (inference):

  • Record the month with a lifelog.
  • Have the person label episodes each evening.
  • Run recall sessions cued by the person's own images, early and late, under fMRI.

  The lifelog makes accuracy of recall scoreable per memory: a candidate held-out target for the decisive test in architecture section "Deliberative recall under recording".

• E3 (verified, availability not checked): the paper offers the data through the author's institutional website, not a repository. The author has since moved institutions, so availability needs checking before E3 relies on it.