Position Paper
What Deserves to Persist
A position paper on earned memory, ephemeral software, and the digital self.
Memory should be earned, not assumed. Persistence should be proven, not granted by default.
The paper starts from the practical frustration that created SR-SI: an AI agent waking up fluent and blank, capable in the moment but unable to carry durable orientation unless the context survives outside the conversation.
From there, the argument widens. Codebases, design systems, organizations, generated tools, and personal memory all face the same pressure: what should persist, who decides, and what should be allowed to decay.
The governing idea
Accumulation is not memory. A transcript can store more, a database can retrieve more, and a notes app can preserve more, while none of them answer what should matter later.
Earned memory asks for judgment before permanence. A useful system should preserve the context that helps future action, downgrade what no longer deserves confidence, and let some material disappear.
Why this follows SR-SI
SR-SI proved the first version of the pattern in software work. A system does not need to store everything to behave coherently over time. It needs a compressed, maintained, consulted, and governed record.
That makes SR-SI more than an AI workflow method. It becomes a proof point for a wider memory discipline: preserve only what helps the next reconstruction.
Four requirements of earned memory
Provenance
The system needs to know whether a claim came from code, a decision record, a user correction, or a stale assumption.
Compression
The record has to stay shallow enough to guide retrieval without becoming another context burden.
Decay
Some memory should expire, some should be downgraded, and some should be deleted before it becomes confidently wrong.
Governance
The agent can maintain the machinery, but the human has to own the policy for what persists.
The systems it touches
The paper is not a product announcement. It is a conceptual ladder that follows one memory discipline across different substrates.
- 01 SR-SI proves that AI context can be reconstructed from a maintained external record.
- 02 Design-system memory turns visual intent into build intelligence for future interfaces.
- 03 Organizational memory lets teams preserve intent, constraints, and tradeoffs across decisions.
- 04 Ephemeral software lets tools appear for a need and earn permanence only through use.
- 05 Personal memory raises the harder question of which Anchor Points should shape future reconstruction.
- 06 Self-indexing systems test the threshold where procedural continuity becomes operationally serious.
The digital-self threshold
The most delicate part of the argument is the personal layer. If memory is not storage, then personal memory systems should not preserve every passing mood at equal weight.
The paper introduces Anchor Points as a cautious way to think about this: moments, decisions, ruptures, commitments, and patterns that change the reconstruction of a life. It also asks what happens when a reflective system tracks its own influence on the person it claims to remember.
Related essay sequence
Next step
Stop guessing. Move to execution.