Persistent AI memory faces a problem that ordinary retrieval benchmarks can conceal. An association may become easier to recall after repeated exposure, while also becoming harder to replace when it is no longer correct.
Velorin Intelligence investigated that tension through a sequence of retrieval comparisons, controlled model-answer evaluations, versioned-memory tests, and restart-recovery experiments. The studies used synthetic fictional identities and, for some early comparisons, an externally authored fictional dialogue benchmark.
The principal negative finding was that the tested neural memory adapter did not establish an advantage over simpler source-aware retrieval.
In a 29-question matched answer-level comparison, the adapter and hybrid lexical–semantic retrieval each produced 14 source-grounded successful answers. The study was small, used a model-distinct rather than human reviewer, and flagged source-reference issues in 15 questions. It does not justify a neural-learning superiority claim.
A separate structured experiment then found a more practical failure. After repeated reinforcement, an online associative trace correctly replaced obsolete facts in only 6 of 20 correction cases. A source-backed ledger retained the current answer reliably in the tested conditions.
That result changed the design question. Rather than increasing reinforcement, the follow-up experiments separated the strength of an association from the authority of an explicit correction.
A versioned memory trace with correction precedence and deletion tombstones passed its own controlled synthetic tests. Subsequent frozen-model experiments parsed fictional conversations into memory events, and an offline persistence study recovered interrupted transactions across real process restarts.
These results are promising for versioned, source-grounded memory, but they are not evidence that a language model acquired durable experience through changes to its own weights.
Findings and limits
Retrieval remained a strong baseline. Learning-based retrieval did not outperform the tested hybrid retrieval on the matched, source-grounded answer comparison.
Repetition alone was unsafe for correction. Reinforced obsolete facts could persist; explicit revision precedence corrected that problem in structured follow-up tests.
Source identity and speaker intent are different problems. An authorization gate can reject an untrusted message, but cannot fix an authenticated owner’s quoted request that has been misunderstood.
Forgetting has a storage boundary. The persistence experiment removed erased facts from its registered active stores and backup, while deliberately unregistered older snapshots retained them.
Offline compatibility is not production readiness. The final test exercised an existing assistant system’s governed-memory reading interface on correctly prepared synthetic evidence. It did not test a live assistant deployment or real conversations.
Research disposition
Status: Completed and parked, November 2026.
The findings support an evidence-first design in which source provenance, explicit revision authority, and controlled deletion are separate from retrieval salience.
They do not establish general reliability, production-grade security, a neural-weight-learning advantage, or readiness for deployment in an assistant.
The original technical protocols, unfavorable results, model-output records, and scoped execution receipts have been preserved privately. Because some source materials and evaluation artifacts are not published, this is not an independently reproducible public artifact release. No external peer review is claimed.
Further research should begin only with a specific failure to explain or an independently scoped, write-disabled deployment-qualification trial.
Velorin Intelligence · Persistent Systems / Reliability · Evidence cutoff: November 5, 2026