We gave personality control of memory. It cost 31 points.
The most seductive idea we shipped this year was wrong, and we have the receipts.
Mnemosyne OS has a cognitive lens: sixteen MBTI-style archetypes you switch from the chat rail, so the assistant thinks as an INTJ or an ENFP for this one question. The obvious implementation is prompt-side, a persona line in the system prompt. We went further, because the further version sounded so right:
An archetype that keeps its options open (ENFP) should read a wider, more diverse slice of your memory than one that converges (ISTJ).
So the lens also steered the retrieval layer: each archetype mapped onto a memory-selection profile, a diversity weight (MMR λ) and a selection strategy, from plain top-k to MMR to a low-discrepancy sampler we nickname phyllotaxis. Personality wouldn’t just change how the assistant speaks. It would change what it remembers.
If you build memory systems, you have felt the pull of this idea. Here is what it measures like.
The measurement
A 61-question in-vault slice of LongMemEval, factual recall over long-term conversational memory. Identical embeddings across all conditions; the lens only re-ranks and selects among already-retrieved candidates. One control, several lens profiles, physically executed runs.
| Condition | Accuracy |
|---|---|
| No lens (control) | 75.4% |
| Converging profiles (ISTJ-like) | 75.4%, identical verdicts |
| Maximally divergent (λ=0.40 + low-discrepancy) | 44.3% |
Minus 31.1 points. Paired analysis: 21 control hits lost, 2 gained, one-sided sign test p < 10⁻⁴. And the direction of the design was refuted, not just its magnitude: the divergent profile was supposed to shine on multi-session questions. It dropped them from 3/10 to 1/10, and crushed temporal reasoning from 9/10 to 2/10.
The ablation, two mild levers, one catastrophe
| Configuration | Cost vs control |
|---|---|
| Diversity weight alone (λ=0.40, MMR kept) | −11.5 pts |
| Low-discrepancy sampling alone (λ kept) | −6.5 pts |
| Both together | −31.1 pts |
The two levers are individually survivable and jointly catastrophic, a super-additive dilution of the relevance head. Each one weakens the grip on the right memory; together they let go of it.
Two more facts we owe you. A full control replicate reproduced all 61 per-question verdicts exactly, run-to-run noise on this rig is zero, so the gap is the gap. And a design confession: because our archetype→retrieval map derived from two binary axes, the sixteen personalities collapsed into four retrieval behaviours. Sixteen faces, four memories.
The bug we didn’t see coming
The style side had its own lesson. Tone lines are accuracy-free, but our persona vocabulary was written in French, and a French tone line inside an otherwise-English prompt turned out to be a language instruction: lens arms answered an English benchmark in French in 8 to 54% of cases, whatever the rules below said. A persona block is never just flavour text. The model reads all of it.
What shipped instead
In v1.3.6 the lens is a costume, not a filter. It changes one thing, the voice: tone, angle, manner. Retrieval stays exact, whoever is speaking. The measured map stays in the codebase as the record of why, and re-coupling personality to retrieval now requires a measured task-type guard first, not a hunch, however elegant.
The general lesson, stated plainly: personality-conditioned prompting is roughly free; personality-conditioned retrieval carries a measurable factual cost. If your product promises memory, the memory must not depend on the mood you dressed the assistant in.
Why publish a negative result
Because the claim “our memory is exact” is only worth what you can check. The 72.9% headline on our benchmark page ships with an audit kit; this experiment ships with its numbers, its refuted hypothesis and its bug. Mnemosyne OS is a local-first memory OS, the whole point is that you can trust what it remembers. That trust is built with measurements, including the ones that told us no.
The cognitive lens, costume edition, is live in Mnemosyne OS v1.3.6. The architecture behind the memory engine is described in the Resonance Engine whitepaper.