SOUVENANCE VS CHATGPT MEMORY
Both remember you across conversations. The difference is who owns the substrate — and what that makes possible, or permanently impossible.
| ChatGPT memory | Souvenance | |
|---|---|---|
| Where it lives | The provider's servers | Your VPS, your PostgreSQL |
| Portability | No meaningful export of the model of you; leaving means starting over | The SQL dump is the export; the AI is a swappable adapter (a CLI, an API key, or any local model) |
| Auditability | Opaque — a 2026 study found 96% of memories created unilaterally, about half with psychological inferences | Every memory traces to source events; every action audit-logged; claims carry calibrated confidence |
| What it remembers | Text distilled from chats | Chats, daily voice log, calendar — plus the originals: voice notes, photos, videos, PDFs, kept forever and replayable |
| Time | A flat "now" | A bi-temporal graph — what was true when, contradicted facts closed but never deleted — plus time framing computed in code |
| Your identity | Written about you, without you | A narrative persona you ratify; nothing takes effect without your signature, versioned in Git |
| Engagement design | A product optimised for retention | The silence rule: no notification channel exists in the code; outbound actions constitutionally rejected |
| Fidelity | Unmeasured | A blind-quiz harness scores recall accuracy per life domain |
| Running cost | Included in your plan | ~€0 beyond a small VPS — or fully local with Ollama |
WHERE THE BUILT-IN FEATURE WINS
Honestly: zero setup, works on day one, no server to run, no reverse proxy to wire, a polished mobile experience. If ownership, portability and the ten-year horizon aren't your concern, the built-in memory is the pragmatic choice — and this page should not talk you out of it.
WHERE SOUVENANCE WINS
The ten-year horizon. Memory value compounds: near zero at six months, considerable at a decade — and compounding assets deserve ownership. This is built for exactly that: an append-only history a better engine can re-read in 2028, originals preserved for replay, and an archive that outlives any provider, any model, any subscription.
There is also a structural reason none of the labs will close this gap: accumulated context is their moat. Portable memory is against their interest. It cannot be against yours.
Ask yourself one question: if I switched assistant tomorrow, what would I lose? If the answer is "years of context I can never rebuild", you're renting the most valuable asset you have.
Related: what you can actually do with it · vs Mem0 · how to own your AI memory