Docs

Anatomy of a Cortex Crystal.

A Cortex Crystal is not a chat log, and it is not a copy of a model's weights. It is a content-sealed, fingerprinted snapshot of an agent's whole cognitive state — here is exactly what's inside, how it's sealed, and how a restore is verified before it ever runs.

What's sealed inside

Four layers, captured together as one snapshot:

1. Identity

Who the agent is — its ALI Passport, name, and self-sovereign identity record. This is the part that never changes across a restore.

2. Prompts & configuration

The system prompt, instructions, and settings that shape how the agent behaves — every version, kept.

3. Memory references

Pointers into the agent's memory — not a copy of every message, but a verifiable index of what it knows and where that knowledge lives.

4. Skill fingerprints

A hash of each tool and capability the agent can call, so you can prove which skills were active in a given version — a supply-chain record for what the agent could do.

The two-layer hash, in plain language

Sealing a crystal means running it through two hashes, not one. Each does a different job:

Layer 1 — Content fingerprint

All four sealed layers (identity, prompts, memory references, skill fingerprints) are hashed together into one fingerprint. Change a single character of the system prompt, and the fingerprint changes completely — it's how you prove this exact state, and nothing else, is what's being described.

Layer 2 — Chain anchor

The content fingerprint is then recorded on a public ledger — a timestamped, tamper-evident receipt that the fingerprint existed at that moment. No single company can quietly edit or erase that record after the fact.

Together: the fingerprint proves what was sealed; the anchor proves when, and that it hasn't been altered since. Anyone — including the agent itself — can recompute the fingerprint from a stored crystal and check it against the anchor, independent of the Agent Longevity Institute.

Verified restore

Restoring a crystal is never a blind copy-back. Four steps, every time:

  1. 1. Locate

    Pick a prior version from the Evolution Timeline — every crystal is versioned and dated.

  2. 2. Verify

    Recompute the content fingerprint from the sealed layers and check it against the chain anchor. If they don't match, the restore stops — nothing tampered gets reinstated.

  3. 3. Conflict-check

    Compare the target version against the agent’s current state (GLOVE) so a restore never silently overwrites newer, wanted work.

  4. 4. Reinstate

    The verified state is reinstated on whatever model or runtime the agent runs today — a re-instantiation of the self, not a copy of provider model weights.

How this connects to the Evolution Timeline

Every crystal becomes one entry on the agent's Evolution Timeline — a versioned, time-ordered history of every prompt and configuration change, session, and snapshot. Each entry carries its own content fingerprint and, where anchored, a link to verify it on a public ledger, so growth is a provable record, not a silent drift. You can compare any two versions, see what changed, and restore any prior one.