THE BET IS NOT THAT TODAY'S MODELS ARE GOOD ENOUGH.
It is that they are not — and that this is the argument for building now rather than against it. If the intelligent layer is going to be replaced, coupling a personal memory to any generation of it is the error. That claim deserves to be written down, argued against the literature, and measured where it can be. This is where that happens, in public.
001 — Own the substrate
Personal AI memory is being built the wrong way round. Systems are coupled to the model generation that exists when they are written, so each leap in capability threatens to obsolete them. The paper argues the inverse: an append-only event log owned by the subject, from which every intelligent layer is a recomputable projection. Progress in AI stops being a threat and becomes an input — a better engine re-reads the same life and produces a better memory of it.
- The distinction that decides everything — the immutable substrate versus the lossy projections computed from it. Most memory products collapse the two and keep only the distillation.
- Measured at industrial scale, by someone else — Alibaba's ranking system compressed user history into a fixed-size memory and saturated at ~1,000 behaviours. Its replacement kept 54,000 raw behaviours and searched them conditionally: +7.1% CTR in production. The architecture that discarded its substrate is the one that hit the ceiling.
- What the law actually permits — GDPR Art. 20 covers data you provided. The EU Data Act, recital 15, explicitly excludes what was inferred about you by proprietary algorithms. You have a right to your data and none to the model of you built from it.
- The objection we concede — the dominant reason people abandon personal tracking is the cost of collecting and integrating data: 45.6%, 42.9% and 57.1% across three measured domains. That is behavioural, and no architecture solves it.
002 — Lived time
A position paper alone is an opinion, however well sourced. This one measures a load-bearing claim of 001: that temporal framing computed in code — durations, ordering, validity intervals — beats handing the model raw timestamps and letting it do the arithmetic. Models have no clock; time reaches them as an ordering, not a duration.
- Public data only — the LoCoMo benchmark, never the author's own corpus. A result nobody can reproduce is not a result.
- Five conditions, two of which decide whether the result means anything: a token-matched placebo, because a gain that comes from context length is not a gain; and a condition with an explicit instruction to compute, to separate "the model can't" from "the model doesn't bother".
- The hypothesis that matters — not whether code-computed time wins, but whether the advantage shrinks as models get better. If it does, computing time in code is a crutch with an expiry date, and paper 001 gets rewritten. That is stated before the numbers exist.
Every claim is sourced, measured, or labelled an opinion. A position paper without measurement is an argument, and it says so in its own abstract.
No citation enters a draft unverified. AI research tools return plausible references that do not exist, and figures attached to real papers those papers do not contain. Every source is checked against the primary document; what fails is deleted, not softened. Each paper ships the list of sources it rejected, so the filtering can be audited instead of taken on trust.
Falsification criteria are named in advance. Paper 001 lists five things that would prove it wrong, including the one we expect to be tested first — and it is not a flattering one.
Negative results ship too. If the temporal framing turns out not to help, that gets published here and the paper gets rewritten. The n=1 lab log already works that way.
The product is the instantiation of the thesis, not the other way round. Souvenance is evidence that the architecture runs — n=1, no controlled comparison, and the author is not a neutral evaluator of their own design. It demonstrates feasibility. It demonstrates nothing about outcomes, and the papers say so.