Inherited Memory Helped My Flies Until the World Changed. Then a Blank Slate Beat Them.

Earlier in this series I killed a brain and the memory lived through it.

Then I dropped that memory into a fresh brain, and the fresh brain woke up believing a life it never lived.

Creepy. Useful. One brain, one lifetime.

Here is the part that actually kept me up.

What if the memory never stops getting inherited? What if it gets handed down for generations, and then the world it learned turns out to be a lie?

That is the whole promise of memory-as-inheritance sitting right there. Anything you can pass down, you can pass down for too long.

So I built a family.

The memory that would not die

Same fruit fly circuit as the last few posts. A cue comes in. The fly approaches or avoids. Good outcomes and bad outcomes bend the connection strengths. That is a life.

Then the fly breeds, the body dies, and exactly one thing crosses into the next generation: the external memory graph.

Not the synapses. The synapses rot with the corpse.

The graph is the will.

And there is no forgetting in the will. No decay. No cleanup. Every generation stacks its own experience on top of everything its ancestors already swore was true, and the pile only ever grows.

In a world that stays put, this looks like a dynasty.

Generation two already knows what took generation one an entire life to learn. Generation five is basically preloaded. Watch only the behavior and you would think you were breeding geniuses.

But I logged what was happening inside the memory, not just what the fly did. And inside, the thing was quietly losing its mind.

The behavior sat pinned at the top, calm and correct. Meanwhile the inherited conviction kept climbing with no ceiling, generation after generation, more and more certain about a world that had never once contradicted it.

A stable world hides that completely.

Everybody looks like a prodigy right up until the ground moves.

Which is the only question worth asking.

What happens to a lineage carrying ten generations of absolute certainty when the certainty stops being true?

Give it a world that can actually kill it

A reversal in a spreadsheet is free. I wanted the memory to cost something real, so I stopped hand-feeding the fly and gave it an ecology.

Now a life has stakes. There is food and there is harm. Energy drains. There is age, maturity, and having kids before you die. Read your world well and you eat, survive, and breed. Read it badly and you starve or get killed before you reproduce, and that whole branch of the family is over.

Then I built two worlds that disagree about one thing.

World B (easy):   cue X is usually food     approach X, you eat
World A (harsh):  cue X is usually death    approach X, you die

A second cue stays safe in both. The only thing that flips between the worlds is what the learned cue X means.

So I could raise a family in one world until the inheritance ran deep, then drop a descendant into the other world and just watch. Does the ancestral memory help the newcomer read a strange world, or is it a suicide note written by dead flies who lived somewhere else?

To keep myself honest I always ran a shadow. For every memory-carrying migrant, a matched blank-slate lineage ran in the same world, same conditions, knowing nothing at all.

The blank slate is the whole point. The question is never "did the migrant survive." The question is:

did inheriting anything beat inheriting nothing?

The number I care about is how long the family line keeps going. Call it L. It counts how many generations the lineage keeps successfully breeding across a forty generation horizon. Migrant minus blank slate. Positive means the memory bought you extra generations. Negative means the memory cost you generations you would have had with an empty head.

Then I locked it. The seeds that decide the result, seeds 4001 to 4576, got sealed and frozen before a single one was opened. No peeking. No tuning. No quiet second run if I hated the answer.

One shot.

What the sealed seeds actually said

I ran a full connectome fly against a matched blank slate in both directions, then opened the seeds once.

B to A. The dynasty walks into hell.

This is the fly whose entire family history screamed "approach X, X is food," strolling into a world where X is now death.

L (migrant minus blank slate):  -10.78 generations
p < 0.001   (0 of 10,000 resamples crossed zero)
never recovered inside 40 generations

The inherited memory did not cost them a little. It cost the line about eleven generations of survival compared to knowing nothing at all.

Sit with that. The blank slate, the fly that inherited absolutely nothing, outlived the fly that inherited ten generations of hard-won family wisdom. And across the entire forty generation horizon the poisoned lineage never once clawed back level. A single descendant, living a single honest life, getting bitten by the truth over and over, could not shout down ten dead ancestors insisting the old thing was safe.

That is the result I came for. Inherited memory turning into a liability so heavy that ignorance wins. Not a metaphor. A measured, sealed, eleven generation grave.

And then the experiment looked at me and refused to give me the clean version.

A to B. Hell's memory walks into paradise.

I wanted a mirror. Caution learned in a lethal world should be dead weight in a safe one, so the migrant should eat dirt again.

It did not.

L (migrant minus blank slate):  +3.75 generations
p = 0.088   (not significant)
unresolved

The estimate even leans slightly positive, and it does not reach significance in either direction. In the easy world the safe cue keeps almost everyone alive anyway, so being wrong about X barely costs a thing. A forgiving world just does not squeeze hard enough for the stale belief to matter.

So this direction did not resolve. Not zero. Not "invisible." Just genuinely unsettled.

There was one more knife in this direction. Raising a family in the harsh world at all is savage. Two thirds of the lineages I tried to grow there died out before they could even finish handing down their memory, so only 32 of them survived to be tested. The harsh world does not just punish bad beliefs. It punishes existing.

The honest shape of it

I could have written this up as a tidy law. Memory helps at home, memory hurts abroad, symmetric, clean, roll credits. That would have been a lie, and the sealed seeds would have known.

The real shape is lopsided. When the world turned hostile, inherited memory was badly and permanently harmful, bad enough that a blank slate beat it. When the world was kind, the inheritance mostly shrugged and the numbers came back a maybe.

The asymmetry is the finding. Not a smudge on it.

And it lands exactly where I have been pointing this whole series:

Useful memory should lower the cost of learning what is still true without raising the cost of discovering what is no longer true.

The lineage aced the first half and face-planted the second. It got competent for free. But the second the ground moved, all that inherited competence hardened into a stubborn wrong prior it paid for in dead children. Nothing anywhere in the memory knew how to say "that was true then, and it is not true now."

This is not a fly problem.

This is every system that carries memory forward and cannot tell stale from valid. Give an agent long-term memory and it inherits this exact failure the instant its world drifts out from under it. The better the old memory was, the harder it fights the truth.

I am not dragging the flies into a genuinely new world yet. This result gets to stand on its own first. It is the clean end of one question: what inherited memory does when the world it learned changes underneath it.

Next: what happens when the world is not a reversal but something the ancestors never saw at all, and whether old memory is a head start or a lie when the present stops looking like the past.

Clone it and break it.

The toy repo from Arc I is still here:

git clone https://github.com/constant-itis/flymem && cd flymem && python3 flymem.py

The mushroom body wiring under the real-fly work is public too, so you can pull the exact same connectome and check whether I read it right:

pip install neuprint-python
# point a Client at server "neuprint.janelia.org", dataset "male-cns:v1.0"
# (free per-user token from the site), then fetch the KC / MBON / DAN populations