Knowledge your agents can cite, check and trust
Mnemontron is knowledge infrastructure that keeps the original evidence behind every answer. It stores sources as content-addressed, versioned evidence, links each claim to the exact spans that support or oppose it, and returns citations that an agent or a person can check.
Status: pre-launch. Mnemontron is in development. There is no public API, no hosted service and no pricing yet. The capabilities below are implemented and tested in our own development environment; they have not been independently verified or validated in production. Early access is by request.
Request early accessSee the measured results
What it is
Mnemontron is a system for storing, organizing and retrieving knowledge so that every answer can be traced back to its source. Instead of returning an unattributed summary, it returns claims with citations to specific evidence: which file, which version, which span, when it was fetched and who holds custody of it.
Who it is for
- Teams building AI agents that need answers grounded in sources they can show, and need to know when the evidence is missing, stale or contested.
- Research, compliance and knowledge teams that need a record of where each fact came from, who verified it and how it changed over time.
- Organizations with strict data boundaries that need tenant isolation, encryption at rest, passkey-only sign-in and an audit trail.
How it works
- Preserve the evidence. Each source is stored by its BLAKE2b content hash with its location, retrieval time, media type, version, license and custody status. Originals, extracts, embeddings and claims are kept distinct.
- Extract attributed claims. Claims link to the exact spans that support or oppose them, with an epistemic status. Verification by an agent (
agent_verified) is recorded separately from verification by a person, together with who verified it. - Retrieve with citations. Hybrid retrieval combines BM25 lexical search, dense embeddings and reranking. Every result carries citations to evidence hashes and versions, and the system declines to answer when the evidence is insufficient.
- Keep it honest over time. Contradictions, corrections and retractions are tracked, historical questions can be answered as of a past date, and results flag sources that are stale or missing without silently changing the ranking.
Capabilities
- Content-addressed evidence with custody. Immutable, versioned evidence identified by BLAKE2b hash, with lineage from originals to derived material and four custody states.
- Rights-aware retrieval. Revoking a source hides it from retrieval, purges it from search and vector indexes and caches, and blocks re-ingest.
- Signed receipts. Query and citation receipts are signed with Ed25519 and carry a nonce and expiry; a receipt presented a second time is flagged as a replay.
- Built for agents. A Model Context Protocol (MCP) server, an A2A agent card, and
llms.txtsupport are part of the product. They are not publicly available yet. - Security by default. Passkey-only sign-in, step-up re-authentication, per-tenant AES-256-GCM envelope encryption, row-level tenant isolation and a hash-chained audit log. See Security.
- Standing red-team suite. Automated attack tests for prompt injection in ingested pages, compromised tool outputs, poisoned federation imports, receipt forgery and replay, and cross-tenant leakage.
Measured so far
These are internal benchmarks run by the Mnemontron team, not independent results. Details and caveats are on the benchmarks page.
| Measurement | Result |
|---|---|
| Natural Questions test (n=3,588), BM25 recall at 20 over 21,015,324 Wikipedia passages | 58.5% |
| Fast-path search latency, 95th percentile, same run | 105 ms |
| Internal held-out set (N=213): answerable questions answered | 90.1% (91 of 101) |
| Internal held-out set (N=213): unanswerable questions correctly declined | 88.4% (99 of 112) |
Contact
Mnemontron is not available yet. We are looking for early design partners who need cited, auditable knowledge for agents or teams.
A public contact address for mnemontron.com has not been published yet. It will be listed here as soon as it is live. Please check back.