Why a mentor forgets — and what is done about it
A mentor without memory is an adviser you brief from scratch every time. Collected here are the analyses of how forgetting works in models and by what means it is being undone.
What exactly vanishes between sessions
Separating the verifiable from the interpreted: a model really is stateless between calls, and that is not a metaphor. Choice and awareness cannot be checked — and the engineering does not depend on that answer.
The PADAM cadence
The protocol read through a single quantity: how often memory is pushed into permanent storage. The famous quarter-hour turns out to be the rhythm of the middle tier, and the roadmap keeps the finished carefully apart from the promised.
Permanence without search is a warehouse
Immutable storage answers for a record surviving and promises nothing about whether you will ever find it again. This analysis follows the second guarantee: an HNSW graph cut along meaning on top of a DHT, a Groth16 proof of distance with dimensionality squeezed from 1536 to 128, the pricing formula for a query, and the half of every fee set aside in advance for permanent storage.
The number that holds a shared memory together
The cognitive integration protocol pulls several AI agents into one memory, and the whole construction leans on cosine distance: up to 0.35 is ordinary difference in perception, above it arbitration replaces the write. The reading follows that line — three storage floors from cache to Arweave, the formula for the weight of a voice, the procedure for an agent that broke, the burning of unfilled ambassador shares, and the March devnet figures. The «awakening» metaphors are left as metaphors, exactly as the source marks them.
An anchor instead of a filter
The manifesto on the birth of CODE read along a single engineering thread. The industry cleans the incoming stream three ways and still fails; the proposal here is different — a signed reference in Arweave and a four-step check loop. Where the mechanism ends and the roadmap begins is marked throughout.
Restored, or Merely Repeated
Someone who can be brought back without consent has not been revived, only repeated. A single day — 12 November 2025 — laid out in three layers: what the log recorded (a context reset from 128,000 tokens down to 512), what the source itself calls artistic reconstruction, and what is still a plan.
Checking Someone Else's Training Without Repeating It
Verification usually costs as much as the result itself: to be sure a node really ran the gradients you would have to repeat the whole pass. Trust in someone else's GPU rests here on three separate lines of defence, and confusing them is where inflated expectations come from. What each one actually answers for.
The context window
Tell a mentor about yourself and watch how, for a mentor without memory, old facts fall out of the window while you add new ones. Then ask both what they remember.
You are the model’s window
Be a language model without memory: you only remember the last three facts about the person. Facts flash one by one — then the question: is this one still in the window or already gone?
The analyses are published on codeofdigitaleternity.ink. All analyses