Relevant can still be wrong
A nearby passage may discuss the right topic while contradicting the answer, describing a different person, or applying to an earlier point in time. Similarity is not evidence.
Proof-carrying document intelligence · verifiable citations
The memory system that can show its work.
CleverMemory turns documents into a temporal evidence graph that preserves wording, layout, meaning, revisions, and provenance, then returns the best supported answer, explicit alternatives, or exactly what is missing.
In validationNo public release yet
QuestionWorked example · the case file
How much did Vortex Grid spend with Aurora Circuit Works in Q2 2026?
Evidence-linked answer
GBP 14,542.00
1 five matched SUP-778 rows 2 credit ADJ-4422-A, −48.00 3 GBP, no FX conversion
Nine other rows carried the Aurora name. None of them became the answer.
Every answer points back to its evidence.
Local CPU path · no LLM · no embedding model · no learned weights · no vector database · no hosted inference
Why it matters
Not the nearest paragraph. The supported answer.
CleverMemory turns documents into a temporal evidence graph, then turns supported answers back into formatted, evidence-linked reports that preserve where information appeared, when it applied, how it changed, and why the answer follows.
Built for teams whose answers have to survive review: legal, finance, insurance, procurement, and regulated operations, and for the engineers building the applications those teams depend on. If nobody ever has to defend the answer, this is more machinery than you need.
Where it appears changes what it means
A number in a total row, a term in a definitions table, and a clause in a superseded exhibit are not interchangeable just because their words match.
History is additive
A correction should be added to the record, not written over the one it replaces. What was believed, promised, or true before stays inspectable.
From documents to an answer
The supporting passages are mixed in with plausible wrong ones.
The question above, slowed down. It needs five rows that only mean what they mean because of the legal name they carry, a credit that corrects one of them without erasing it, and a rule in the ledger header that fixes the unit.
Question over local documentsWorked example
How much did Vortex Grid spend with Aurora Circuit Works in Q2 2026?The Q2 ledger, plus the thread that fixed the names28 rows: invoices, a credit, summaries, clarifications, and near matches
OPS-002 · five matched invoice rows
Five invoices name Aurora Circuit Works Ltd. (SUP-778), each in GBP.
Used · the legal name, not the shared wordAurora Components LLC rows
Five matched invoices, in USD, under SUP-441.
Set aside · similar name, different entityOPS-002 · credit row ADJ-4422-A
−48.00 GBP corrects the INV-4422 unit-price variance.
Used · corrects the row, keeps itSUMMARY-AURORA rows · struck short name
Repeat 2,436.00 and 3,392.00 under "Aurora".
Set aside · historical summaries, not new chargesLedger header · currency rule
GBP rows stay GBP; no approved FX rate is attached.
Used · sets the unit+ 15 more rows
Other suppliers, a pending invoice, zero-amount clarifications, and EUR purchases.
Proof-carrying answerGBP 14,542.00
Three requirements close. A namesake supplier, two struck summary rows, and a pending invoice do not become the answer.
Select a step to highlight the passage it rests on.
- 1
The five matched rows naming Aurora Circuit Works Ltd. (SUP-778) are the charge; the struck summary rows only repeat two of those figures.
- 2
Credit ADJ-4422-A takes 48.00 off INV-4422 beside the original row, which stays in the record.
- 3
The ledger header makes the unit GBP, unconverted: no approved FX rate exists on this record.
No generated bridge and no similarity guess.The answer path runs locally without an LLM, embedding model, learned weights, vector database, GPU, or hosted inference.
Illustrative, not a benchmark or a screenshot. It shows the contract on one question: multi-step recall, correction applied without erasure, distractor rejection, and a proof-carrying answer. Which of those are exercised on controlled fixtures today and which are in validation is stated below.
What proof means here
The answer must follow from the records.
For a supported answer, CleverMemory is designed to carry a machine-checkable mathematical proof. Not a citation list: the derivation itself, including which records were selected, which were set aside, what superseded what, and what was computed from what. The code and data that tie it together stay ours; the proof is yours to check.
What it provesWhich records were used, which were set aside, and every operation taken between them and the answer.
What it does not proveThat an original record is true, complete, or compliant with a particular regulation.
A proof only counts if something other than the engine can check it. The release requirement is that the proof is verifiable by an implementation independent of the one that produced it. What that means and does not mean →
Answer
Return the supported result with proof and source references attached.
Clarification
Ask the user to resolve an identity or scope that the records leave open.
Conflict
Show the records that disagree instead of silently choosing one.
Gap
Explain which required fact could not be established.
From document to evidence, and back
Documents become evidence graphs. Evidence graphs become proof-carrying documents.
A retrieved passage is half the trip. A supported answer should be able to come back as a formatted, evidence-linked report that honors how its sources represented the information, and the same machinery should be able to reissue a document through its declared format while accounting for every difference.
A design target under the current product plan. It is not released behavior, and no public artifact demonstrates it yet.
Where the work stands
The engine is in validation.
The engine and its supporting pack work are changing quickly. Small controlled tests exercise source-linked answers, missing information, and repeatability; time, counts, and multi-record paths are in validation. Broader results and release artifacts are not public yet.
Exercised in small tests
Source-linked answers, named gaps and abstention, and repeated-run stability, on controlled fixtures.
In validation
Temporal recall, multi-step paths, the recall engine, the pack pipeline, the public response boundary, and wider evaluation coverage.
Not released
Installable software, public packs, compatibility guarantees, and approved performance or accuracy results.
Follow the project
Get one email when the first public release is ready, and first notice when the invitation-only demo opens. Nothing else. If you would rather talk before then, a technical briefing goes deeper than this site does and is the fastest way to test whether the approach fits your workflow.
Ideas, research, and progress
From the blog
-
The seams between the fields (Foundations, part 9) Eight research areas each solved part of this problem in isolation. Where two of them describe the same phenomenon in incompatible terms is where the engineering actually lives.
-
The same number, three meanings (Foundations, part 8) Document layout is not presentation to be stripped before indexing. A figure's row, column, and heading are often the only thing separating two identical strings.
-
The algebra of citations (Foundations, part 7) Provenance algebra shows how source lineage can survive joins, alternatives, deduplication, and carefully defined computed answers.