Glossary

Plain definitions of the terms used across the Mnemontron site and product, each with how Mnemontron applies it. For the product overview, start at the home page.

Provenance-aware knowledge infrastructure

A system for storing and retrieving knowledge that keeps the origin of every fact attached to it. Mnemontron is built this way: each answer carries citations to the preserved source, the version and the exact span it came from.

Evidence chain

The three linked records behind every Mnemontron answer: the preserved source, the attributed claim extracted from it, and a signed receipt that connects the answer to the source's content hash and version. Following the chain from receipt to source is how a reader checks an answer.

Content addressing

Identifying a piece of data by a hash of its bytes instead of by a name or location. Mnemontron identifies evidence by its BLAKE2b hash, so any change to the bytes produces a new identity and a new version.

Custody status

A label that says who holds the original copy of a source. Mnemontron uses four states: locally preserved, contractually preserved by a custodian, reference only, and unavailable or unverified.

Attributed claim

A statement extracted from evidence and linked to the exact spans that support or oppose it, with an epistemic status. Claims, originals, extracts and embeddings are stored as distinct records.

Agent verification and human verification

Two separate records of who checked a claim. A claim checked by an agent is marked agent_verified; a claim checked by a person is marked human-verified with who verified it. Mnemontron never merges the two.

Abstention

Declining to answer when the available evidence does not support an answer. Mnemontron measures abstention as its own result: how often unanswerable questions are correctly declined.

Answer-support check

A test of whether the cited evidence actually supports an answer, run separately from the test of whether the evidence is relevant to the question.

Hybrid retrieval

Combining several retrieval methods and merging their results. Mnemontron combines BM25 lexical search, dense embeddings and reranking, and every result carries citations to evidence hashes and versions.

Signed receipt

A digitally signed record that links an answer to its evidence. Mnemontron receipts are Ed25519-signed JSON Web Signatures listing each citation's evidence id, content hash, version and URL, with a nonce and an expiry. See how to verify a receipt.

Replay detection

Recognizing a receipt that is presented more than once. Each Mnemontron receipt carries a unique nonce; the verification API records it on first use and flags any later presentation as a replay.

Rights revocation

Withdrawing permission to use a source. When a source is revoked in Mnemontron, it is hidden from retrieval, purged from search and vector indexes and caches, and blocked from being ingested again.

Historical retrieval

Answering a question as of a past date. Mnemontron keeps versions of evidence and claims so that a query can be run against what was known at that time, and versions can be compared.

Staleness and missing-material flags

Markers on results whose sources are out of date or no longer available. Mnemontron shows these flags next to results without silently changing the ranking.

Envelope encryption

Encrypting data with a data key and then encrypting that data key with a master key. Mnemontron seals each tenant's evidence with its own AES-256-GCM data key, wrapped by a master key.

Row-level tenant isolation

Restricting every database query to the rows of one tenant. Mnemontron uses Postgres row-level security, on top of membership checks, so one tenant's queries cannot read another tenant's records.

Hash-chained audit log

An audit log in which each entry includes a hash of the previous entry, so deleting or editing an entry breaks the chain. Mnemontron also signs the chain head periodically and writes it to write-once storage.

Capacity projection

An estimate of how a system would behave at a scale that has not been tested. Mnemontron's 100+ exabyte figure is a capacity projection built from measured runs and architectural assumptions; it is not a test that was run. Measured runs are on the benchmarks page.