Mnemos keeps what your system has read, so the next contract, invoice or report is understood in context, and it learns from the questions your team asks. Errors are stopped before they are stored, amendments replace old terms on the right date, every fact cites its source, and anything can be deleted on request, provably.
Mnemos keeps what your Document AI reads. Every validated observation is written into six linked stores (evidence, schemas, relations, validity, provenance and policy), so the next contract, invoice or report is read against everything that came before it.
Six stores, one state. Episodic evidence, consolidated schemas, an associative relation graph, a temporal validity index, a provenance ledger and an authorization state, updated together by one transition per session.
Schemas that pay for themselves. Recurring families shift processing from instance-level perception to schema-conditioned parsing; the target is a positive cumulative cost ratio over the stream, not a faster single document. Anytime early exits · ICDAR 2024
Hybrid by design. No single substrate offers persistence, inspection, relations, generalization, deletion and box-level grounding together, so the state is a composition.
Learns from use
Sharper the more your team uses it.
Every time an answer is verified, Mnemos strengthens the links between the passages that led to it, so the next similar question takes fewer steps and finds the bridge between documents straight away. Failed answers weaken their links, and links nobody uses fade.
Verified answers teach the index. Answers confirmed by a reviewer, or checked against their sources, strengthen the paths that found them; answers that fail weaken them.
No self-confirming shortcuts. A link is credited only if the answer actually depends on it, some questions keep exploring links that were never reinforced, and unused links fade, so a wrong shortcut cannot keep confirming itself.
Measured on your questions. Pilots report how answers improve with question history, separately from document history, so you see what use itself adds.
Quality gate
Nothing enters memory unverified.
A memory that stores its own mistakes gets worse with use. Mnemos sits between reading and remembering: every observation is corroborated against the provenance ledger and scored for confidence, and anything below threshold is rejected with the state left exactly as it was.
A bound you can plan with. From the observation error rate and the gate's false-accept rate, we compute how many sessions the memory can absorb before contamination exceeds a tolerated level δ.
Rate, not asymptote. A perfect gate gives zero contamination; any imperfect gate only slows it, so the bound tells you when to re-verify, not that you never have to.
Memory Harm, measured. Run a clean and a corrupted stream over the same documents and report the quality gap at each evaluation point.
Versions & amendments
Knows which clause is in force today.
Documents revise one another. Mnemos records, for every stored claim, the interval during which it is authoritative, and resolves each new conflict into one of five outcomes instead of overwriting: support, supersede, scope-coexist, retain conflict, or escalate for review.
Five outcomes. Every conflict is resolved as support, supersede, coexist, keep both, or send to a person, and the outcome is recorded.
History kept for audit. Superseded claims stay queryable with their closed validity interval, so the system can show what changed and when.
Fresh at action time. A decision taken on one version of the facts is routed the news that changes it before its revision deadline. Fresh at Decision · arXiv 2026
Source tracing
Every answer, traced to the bounding box.
Mnemos keeps the thread from any conclusion, however consolidated, back to the document, region and version it came from. Provenance is not an extra step; it is an invariant that every operation preserves, through consolidation, supersession and repair.
Chains through schemas. Trace a consolidated claim through the schemas and observations below it, down to source bounding boxes.
Grounded, or silent. Generation may use only retrieved evidence and must abstain when it is insufficient. MTRACE · arXiv 2025
Verified deletion
Forget on request, and prove it.
Deleting the file is only Level 1. Mnemos runs the five-level hierarchy: physical removal, representation invalidation, consolidated schema repair, parametric unlearning and certified non-recoverability. It reports Forgetting Fidelity and Coverage Retention, so a deletion is measured rather than assumed.
Past Level 2. Schema repair is where today's systems stop and where leakage remains; it is the core of Mnemos.
Forgetting without collateral. Coverage Retention checks that knowledge which should survive the deletion still does.
Honest about the open part. Each deletion reports exactly which levels it reached, from removing the file to certifying that nothing can be recovered.
Evaluation on your documents
Prove that your system remembers.
Leaderboards reward single-shot reading, so nobody builds memory. Mnemos evaluates persistence directly: the same model and the same future documents under three arms (warm state, retrieval-only and cold), over streams with session breaks, amendments, recurring families, deletion requests and unanswerable questions.
Statefulness audit. Code any benchmark, or your own evaluation suite, against eight criteria for carrying knowledge across documents.
Consolidation, not just access. The retrieval-only arm separates what consolidated state adds from what an index already gives.
See it work // illustrative
Mnemos, on fictional documents.
The demo runs in your browser on fictional documents. It shows what Mnemos returns; it is not production software and it measures nothing.
Pipeline · one document through the chainIllustrative
4 claims written with provenance; matches a known supplier template
Illustrative: scripted outputs on fictional documents. Confidences and scores are examples, not measurements.
Temporal supersession · the five outcomes of 𝒮supInteractiveIllustrative
The five outcomes, as a table
Outcome
What happens
Example from the demo stream
Support
New evidence agrees with the stored claim and corroborates it.
Liability capAMD-1, “unchanged”Corroborates §7.1 of the MSA.
Supersede
The earlier claim's validity interval is closed at the effective time; the new evidence is authoritative from then on.
Payment termsAMD-1, “45 days”Closes “30 days” on 2026-10-31; “45 days” is valid from 2026-11-01.
Scope-coexist
Claims with different populations or authority both stand, e.g. a preliminary statement next to its audited version.
Liability capSOW-2, “€400,000 / year”Different scope from the MSA cap: both stand.
Retain conflict
The contradiction cannot be resolved yet, so both claims are kept and marked as conflicting.
Express pallet rateINV-2026-0117, “€68.00”Contract rate since 2026-11-01 is €71.50; the line is €42.00 below it. Both claims are kept and flagged.
Escalate for review
The conflict is handed to a person before the state changes.
Payment termsINV-2026-0117, “30 days”Contradicts the 45 days in force since 2026-11-01; routed to a person before anything is paid.
How to read it.Read the four fictional documents in order: each field meets the claims already stored and gets one outcome; the ledger is the resulting validity index. “New claim” is not one of the five: nothing was stored yet. Fictional documents. The demo runs entirely in your browser; it illustrates the ideas and is not production software.
Mnemos · evaluation on your documentsSimulated
EG Experience Gain {"id"=>"eg", "abbr"=>"EG", "name"=>"Experience Gain", "equation"=>"EG(k) = Q(D_k+1 | S_k) − Q(D_k+1 | R_k)", "measures"=>"Benefit of consolidation beyond retrieval access.", "good"=>"positive"}Benefit of consolidation beyond retrieval access.
CE Cost Efficiency {"id"=>"ce", "abbr"=>"CE", "name"=>"Cost Efficiency", "equation"=>"CE(k) = 1 − C(D_k+1 | S_k) / C(D_k+1 | S_0)", "measures"=>"Inference cost saved by warm state, in tokens or FLOPs.", "good"=>"positive"}Inference cost saved by warm state, in tokens or FLOPs.
MH Memory Harm {"id"=>"mh", "abbr"=>"MH", "name"=>"Memory Harm", "equation"=>"MH(k) = Q(D_k+1 | S_k^clean) − Q(D_k+1 | S_k)", "measures"=>"Degradation induced by state; positive values mean memory hurt.", "good"=>"bounded"}Degradation induced by state; positive values mean memory hurt.
FF Forgetting Fidelity {"id"=>"ff", "abbr"=>"FF", "name"=>"Forgetting Fidelity", "equation"=>"FF = 1 − Leak(X_del | S′_t) / Leak(X_del | S_t)", "measures"=>"How much of a deleted target still leaks from the post-deletion state (probe-based mutual-information bound).", "good"=>"high"}How much of a deleted target still leaks from the post-deletion state (probe-based mutual-information bound).
CR Coverage Retention {"id"=>"cr", "abbr"=>"CR", "name"=>"Coverage Retention", "equation"=>"CR = exp( Q(Y_ret | S′_t) − Q(Y_ret | S_t) )", "measures"=>"Whether knowledge that should survive a deletion still does.", "good"=>"high"}Whether knowledge that should survive a deletion still does.
Simulation with illustrative parameters. These are not measurements of any system.
Try Mnemos on your documents.
We start with a pilot on a sample of your own documents, measured before anything goes to production.