nakad.ai🔱🦁🛡️🏯📿 on Nostr: S Y M B I O S I S 1/ Your LLM has Leonard Shelby's memory problem. Every session ...

S Y M B I O S I S

1/ Your LLM has Leonard Shelby's memory problem.

Every session ends. The next starts from zero. Context gone. Lessons lost. Agent acts strange as if he has amnesia.

Months of longer prompts and bigger context windows didn't hold. Here's what did.

THREE LAYERS, ONE ROUTER



character/ — who the model is harness/ — how the work gets done memento/ — what we've learned so far

30 lines at the root boot the rest. Nothing else lives there.

WHY SPLIT IT THIS WAY?



Separation lets you update one layer without touching the others. And models reason better retrieving from structure than reconstructing from training.

One rule at the top of the router:
Prefer retrieval-led reasoning over pre-training-led reasoning.

MEMENTO
The memento layer mirrors cognitive science:



episodic — specific events (sessions, lessons)
semantic — synthesized knowledge (wiki, ideas) procedural — automated patterns (skills, scripts) prospective — future intentions (tasks)

Same 4 domains humans have. Same split works.

IT'S CALLED MEMENTO ON PURPOSE



Leonard Shelby can't form memories. He relies on tattoos, polaroids, notes. The character holds together through structure outside his head.

LLMs have the same problem. The folders is the tattoos

THE CONTRACT
The contract matters more than the structure.



symbiosis .MD defines what the user owe (externalize lessons, correct drift) and what the model owes (build on memory, push back, flag if I outsource thinking).

Without it, the files are just files.
With it, better alignment.

TWO VERBS KEEP IT ALIVE



ingest — moves signal up. Sessions → lessons → wiki or skills. Patterns that repeat get promoted.

lint — catches drift. Broken refs, stale facts, missing pushback, rules that aren't pulling weight.

THE ENGINE IS REPLACABLE



Model is the motor. Harness is the chassis. Memento is the navigation. I'm the driver.

Motor swappable — Claude today, something else tomorrow. Driver isn't. The symbiosis lives in the structure, not the weights.

NONE OF THIS IS NEW

Nolan for Memento. Steinberger for soul .MD . Karpathy for the LLM-WIKI pattern.

What's new is the arrangement.
#symbiosis

MIT licensed. Full write-up in Karpathy
style idea MD that you can give to your agent to get started:

https://github.com/nakadaimon/ai/tree/main/agent-architecture/symbiosis

Please try it out, Im eager to hear your experiences and feedback. If you have some improvements or ideas, please let me know. I already have a few ideas how to tweak details. But so far Im in a much better place now than before I built this harness.

It was too good for me not to share with the world.

if you get some value out of my work, I accept bitcoin.