From Personal Memoirs to History: AI-Powered Testimony Processing

One of our specialties is working with unstructured documentary material: personal memoirs, letters, diaries and recorded testimonies. Here's a real project — a comprehensive documentation of a Jewish family dispersed across four continents during and after the Holocaust.

The Foundation: 40 Years of Accumulated Material

The fourth-generation descendants came to us with:

12 written memoirs — some in Hebrew, some in Yiddish, some in German and English
23 recorded interviews conducted with earlier generations in the 1980s and 90s, often with low audio quality
Hundreds of family letters, never published
Photo albums with handwritten notes — sometimes inaccurate
Official documents — birth certificates, passports, immigration papers

The goal: to build a unified historical narrative of the family, verify facts, bridge contradictions between different memoirs, and surface stories previously known only in fragments.

How Do You Work With "Soft Truth"?

The biggest challenge in working with memoirs: they're subjective. Aunt Rachel remembers the family fleeing Vienna in April 1938; Uncle Jacob is certain it was May. Who's right?

The traditional answer: look for official documents. But what if there are none? What if they contradict? And more importantly — what if the emotional context of a specific memory is actually the more valuable source?

The methodology we developed:

Transcription stage: processing every recorded interview (even those with poor audio) into a full transcript. We used modern AI systems capable of handling background noise and varying accents.
Semantic reading stage: an LLM agent read each text and extracted structure: who? what? where? when? with whom? — every "event" stored in a structured database.
Cross-referencing stage: when two sources discussed the same event, the system compared versions and flagged:
- Details that match and reinforce each other
- Details that contradict — marked with confidence levels
- Unique details from a single source that cannot be verified
Enrichment stage: searching external sources (census records, Jewish organizational archives, Arolsen, USC Shoah Foundation) to verify objective details.

The Transformation: From Collection to Narrative

This is where the real magic happens. Instead of receiving a list of 500 separate events from all sources, we built:

A unified family timeline

An interactive graph showing the life of each family member across 150 years, with the ability to drill into any point and see which sources support it. A grandfather believed to have lived only in Łódź turned out to have regularly traveled to Kraków for work. An emigration thought to be "direct" from Vienna to Palestine turned out to have passed through Italy and Lebanon.

A social network map

The agent identified mentions of over 400 non-family people in the memoirs — friends, neighbors, business partners, rabbis, doctors. 37 of them we were able to identify as known historical figures, which opened a door to complementary research in additional sources.

Resolved contradictions

Rachel was right: the family did leave Vienna in April 1938. Jacob remembered the month the family arrived in Trieste — a month and a half later. Both were telling the truth, each from their own perspective.

Integration Into the Family Story

The final output: a family history book of 380 pages, written by a professional historian — but the scaffolding behind it (the database, the timeline, the cross-verification) was built by AI in a matter of months instead of years.

Every page in the book has a corresponding source file that allows any reader to trace back to the original quote, recording, letter or document underlying every claim.