Memory as a service

Agent memory that cannot be talked into authority

Deterministic recall, injection quarantine, hash-chained provenance, provable deletion — free to use until 1 January 2027.

Try it right now

This runs the real engine on a throwaway database: it writes a sourced fact, writes a hostile memory, shows the hostile one is withheld by default, shows the tombstone, verifies the hash chain, and runs all 33 invariants. Nothing is simulated.

Then keep your own memories

Write a fact and recall it. Your memories are keyed to a visitor identity, not to an account. Send a stable X-CAIN-Visitor header (16–64 hex characters) to keep them across sessions.

Latest measured run (2026-11-02T15:40:06Z): write p50 6.139 ms, recall p50 6.623 ms, invariants 33/33, isolation failures 0.

What it refuses to do

• A memory can never grant authority. Every view says grants_authority: false.
• Text that tries to instruct the model is quarantined and withheld until a named human releases it.
• Delete erases the content but keeps a chained tombstone, so deletion itself is provable.
• A contradiction is flagged for review. It is never silently resolved, and nothing is overwritten.

Read the whole thing

Sixteen short pages in plain English: a two-minute quick start, how saving, searching, quarantine, deleting and the tamper-evident log work, the 33 live checks, connecting your agent, limits, measured numbers and the free preview. Start with the docs →