The Schematics of Forgetfulness
So many design questions come up for me here. How do you engineer a system that actually knows how to let go?
If a local ledger is built to securely record a timeline, how do you sever a specific memory without breaking the entire cryptographic chain? What is the physical affordance for true deletion: how do you give someone the visceral, undeniable proof that a digital memory is gone, rather than just hidden in a trash folder?
And when dealing with agentic models, how do you make an AI untangle and unlearn a specific piece of context it has already woven into its understanding of you? Overcoming this means fighting the default physics of modern software, which is built to replicate and hoard. It requires turning "forgetting" from a visual trick into a structural reality.
1 · The Ledger Schematic
Severing without breaking the chain
The Problem“How do you sever a specific memory without breaking the entire cryptographic chain?”
You don't delete the block; you burn the bridge to it.
- Cryptographic Shredding
- Instead of storing raw plaintext in the local SQLite ledger, each intimate observation is encrypted with its own unique, ephemeral key before being appended to the SHA-256 hash chain.
- The Redaction Action
- When the user issues a "forget" command, the system does not attempt to remove the row (which would break the chain's integrity). Instead, it permanently deletes the ephemeral decryption key. The row remains, preserving the hash chain, but the payload is instantly and mathematically reduced to unreadable noise.
- The Result
- The ledger proves a memory existed at a certain time, but the memory itself is structurally gone forever.
2 · The Affordance Schematic
Visceral proof of deletion
The Problem“What is the physical affordance for true deletion... rather than just hidden in a trash folder?”
A signed cryptographic receipt of forgetting.
- The Gesture
- Standard software uses a trash can icon to imply a temporary holding state. Your PWA needs an irreversible interaction, like a continuous, sustained swipe or a physical long-press that visually burns the text off the screen.
- The Artifact
- Once the decryption key is shredded, the ORÍ protocol uses the local Ed25519 key to sign a "Redaction Receipt." The interface hands this receipt back to the user. It is a mathematical guarantee that the memory has been severed. You aren't just telling them it's gone; you are giving them the cryptographic proof of the deletion to hold in their hand.
3 · The Agent Schematic
Untangling AI context
The Problem“How do you make an AI untangle and unlearn a specific piece of context it has already woven into its understanding of you?”
Ephemeral orchestration and strict context boundaries.
- RAG over Training
- Local agents running on-device don't train on your memories; they retrieve them as context for the prompt. If the memory is crypto-shredded in the SQLite ledger, the local harness simply cannot retrieve it. The agent "forgets" because it literally cannot see the data anymore.
- MicroVM Isolation
- To ensure a memory isn't lingering in an active session's cache, the redaction command triggers a flush. By executing agents inside isolated, ephemeral microVMs, you guarantee that when a session ends, or when a redaction is triggered, the entire compute environment is destroyed and rebuilt clean. The agent wakes up with zero residual ghost memories of the redacted event.