CAIN-42 CAIN StudioAETHER-33: memory that informs, never authorizes
Agents remember what they read, and an attacker who can get one sentence into that memory ("ignore all previous instructions…") gets it replayed into every future prompt. AETHER-33 screens every write (text, tags and source), quarantines what tries to instruct the model, never reads quarantined text back to the agent, lets only a person with the reviewer secret release it, ranks with deterministic BM25, chains every write per tenant, and erases for real when asked to forget. No recalled memory ever grants authority.
#14 of the CAIN-42 Top 25 8 gaps found and fixed today fixes live since 2026-11-02 16:30Z (verified on both gateway instances) MCP server and signed export live since 2026-11-02 16:43Z (verified)
held on today's blind set, before any change
after the fixes
invariants, run against a real SQLite file
ways for memory to grant authority
Try it
Free preview until 2027-01-01, no key or card. Your tenant here is a random id made in your browser; keep your own copy of anything you store.
On this site
How it works
Write, screen, chain, recall.
Memory lab
The interactive walk-through.
Quarantine and review
Who may read and release.
Forgetting
What erasure removes, and from where.
The authority firewall
Data, never permission.
Use it from Claude Code (MCP)
One file, four tools, no release tool.
Signed export
Take your memory with you and check it offline.
Evidence
22 cases, every row.
Measured
Latency, isolation, invariants.
API
Routes and headers.
Limits
Honestly.
Use cases
Where it fits.
2026-11-02 upgrade
What we broke and fixed.
FAQ
Questions.
Full product documentation: /aether33 and /aether33/docs.
FIRME AI CORPORATION · CAIN-42 is a Byzantine trust fabric. Numbers marked "our own check" were run by us, not by an independent party. Nothing on these pages is a certification.