Repeated explanation
Explain changed work criteria again
TENCONIC AI DATA ENGINE
Your latest company knowledge.
Connected to AI answers.
Select the evidence that matters and reflect changing work criteria.
Help your team spend less time explaining context and correcting AI answers.
Company information and work criteria, connected to your existing AI service.
TENCONICAI Data Engine
GPTClaudeLlamaGeminiMistralCompany sourcesWork criteria
EXPERIENCE, CONTINUED.SCROLL TO EXPLORE ↓
01THE PROBLEM
Your information changed.
Who checks the AI’s answer?
Explain the new policy. Review the answer. Correct misapplied criteria.
Even after adopting AI, the repeated work remains.
The refund policy has changed.Which version is this answer based on?Please check the exception again.
Outdated criteria
Old and current criteria appear in the same answer
Repeated review
People recheck the evidence and its application
Problem and scope
Memory, retrieval and update features vary across models and RAG products. Tenconic aims to reduce repeated evidence checks and answer corrections after changes. Customer work-time and operating-cost savings require separate validation.
02STRUCTURE, LEARN & REUSE
Structure company knowledge.
Carry learned decisions into the next task.
Add learning and reuse of work experience
to the definition and connection of knowledge.
Roles and learning capabilities
| Capability | OntologyDefine meaning | Knowledge graphConnect facts | Semantic layerInterpret work | TENCONICLearn & reuse |
|---|---|---|---|---|
| Represent work meanings & relationships | O | O | O | O |
| Reuse defined knowledge & criteria | O | O | O | O |
| Learn work decisions from examples | X | X | X | O |
| Reuse learned decisions after save & restart | X | X | X | O |
Learned decisions reused after save and restart8 fixed synthetic development cases · Decision Cells A & B: 8/8 each · Final drafts: 7/8See the validation scope ↓
Designed for use alongside RAG, Graph RAG and AI agents. Actual integrations and feature availability must be confirmed for each sales version.
Base policyCustomer exceptionChange historyWithdrawn record · Excluded
Applicable evidenceValid evidence with its sources
Comparison basis
This compares basic conceptual roles and learning capabilities, not rankings of competing products. Ontologies, knowledge graphs and semantic layers can be used together.
03HOW IT WORKS
When criteria change,
keep past and current versions distinct.
See the criteria for the relevant point in time, together with their source.
- 01Company sources & work examples
- 02Records & change history
- 03Evidence & supported decisions
- 04Existing AI answers
- 05Review & correction
SYNTHETIC CUSTOMER-SUPPORT RECORDS
Refund-policy change history
SELECTED POLICY RECORD
Criteria recorded on January 21, 2026
Refund request period15 daysSource: policy-05
An example of supplying the same answer and source with less input. Both Tenconic and the comparison RAG passed answer-and-source review on the January 21 question.
04STRUCTURED MEMORY
Learned work decisions,
reused in the next run.
Store and reuse decision criteria acquired from supported examples.
Manage source records separately so the evidence behind an answer can be checked.
Work case
Review
Criteria
Evidence
Correction
Cell
- User intent
- Context history
- Evidence and facts
- Confidence and version
REUSE AFTER SAVE & RESTART
Check for a missing reviewer.
Flag an alternative date when the deadline is close.
Translated extract of the recorded answer: “The reviewer information is missing from the current request.”
“Please review the alternative date (2026-10-30).”
- Decision Cells A & B
- 8/8 each
- Decisions used in answer input
- 8/8
- Final work drafts
- 7/8
Work-learning test conditions and scope
Eight fixed synthetic development cases in S3, 2026-09-14. After a normal save and restart, learned decisions were calculated again and used in the answer input. This was not an independent unseen evaluation; one final draft omitted a supplied reviewer.
Mean request completion time was 333.8 seconds. The test did not isolate the effect of using Cells, and does not establish commercial response-speed or cost advantages, or automatic learning of arbitrary work.
05EVALUATION & EFFICIENCY
Evaluate memory use.
Measure input and response time separately.
Memory QA and the controlled RAG comparison are separate evaluations.
Read each result with its scope and conditions.
MEMORY QA
5 benchmark families · 6 evaluation panels
11,107 questionsLoCoMo · LongMemEval · BEAM · PersonaMem · HaluMem
LoCoMo · Single-hop94.65%796 / 841
LongMemEval · Knowledge update87.18%68 / 78
Aggregated stored scores from a self-evaluation. The two highlighted categories rank first among 15 and 13 systems respectively in the report’s comparison with the 2026-10-08 public snapshot. This was not a simultaneous controlled rerun of competing systems.
Full benchmark results and comparison conditions
| Panel | Questions | Score | Rank / systems |
|---|---|---|---|
| LoCoMo | 1,540 | 87.8571% | 2 / 15 |
| LongMemEval | 500 | 86.6000% | 2 / 13 |
| BEAM 100K | 400 | 0.696769 | 4 / 12 |
| BEAM 10M | 200 | 0.552172 | 7 / 12 |
| PersonaMem | 5,000 | 36.8400% | 4 / 13 |
| HaluMem | 3,467 | 86.4148% | 2 / 13 |
Answers: GPT-4.1-mini-2025-04-14; model-based judging: GPT-4o-mini-2024-07-18. PersonaMem uses multiple-choice correctness. Among the 12 systems with all six panels, Tenconic ranks second on the family-balanced composite (72.5138/100). The composite is not a shared accuracy measure.
Comparison of completed October records v7-20261014 with the OmniMemEval 2026-10-08 public reproduction snapshot. This self-evaluation used a configuration selected through prior exploration; it is distinct from exhaustive human review or independent certification. Answer text is public for 6,100 of the 11,107 stored scored questions.
CONTROLLED RAG COMPARISON
Efficiency measured with the same model and records
71.5%
Fewer input tokens
1,083.81 → 309.25 tokens
39.4%
Shorter answer completion time
897.15 → 543.25 ms
24 synthetic support records · 16 unique questions × 3 runs · Same Qwen3.5-4B · Self-measured against raw-source Chroma K=12
Internal answer-and-source review: Tenconic 15/16 · Comparison 14/16
Efficiency comparison conditions and limitations
Measured 2026-10-05 with Qwen3.5-4B Q4_K_M, the same PC, sources, declared event fields, questions and generation settings. Caller-declared fields—not a natural-language automatic extractor—were provided to every configuration. Five configurations × 16 questions × three runs produced 240 saved answers.
Times are the mean of each question’s three-run median, measured from lookup or retrieval start; ingestion, indexing and model startup are excluded. Tokens are mean actual usage. Output-length differences affect completion time. Review was internal and unblinded; results do not represent all RAG systems or measured customer operating-cost and intervention reductions.
Earlier MAB single-fact integration 64K result
On 2026-10-03, the single-fact integration 64K setting with Ministral-3-8B and the same 8K context showed accuracy of 24% → 51% (about 2.1×) and 66.9% fewer input tokens versus direct recent-record FIFO input. This is not the MAB overall average and is separate from the QA and RAG comparisons above.
06SCALE
Connect to your existing AI service.
Validate one workflow first.
Start with recurring updates and reviews in customer support,
internal guidance or work-request review.
Company information to connect
Policies & rulesCustomer conditionsWork-review examples
TENCONICAI Data Engine
AI builders, operators & enterprise services
- Customer support
- Internal guidance
- Work-request review
API or module integration with existing AI services is the initial approach under consideration. Supported models, inputs, deployment options and SDK availability are confirmed per sales version.
YOUR NEXT STEP
Start with one workflow.
Explore the fit.
Let’s review the recurring updates, checks and sources in your workflow.