Research grounding and novelty boundary
Reviewed November 11, 2026. These are methodological influences, not baselines executed in this release.
- HippoRAG 2 / From RAG to Memory (ICML 2025), official paper: https://proceedings.mlr.press/v267/gutierrez25a.html . The paper combines passage integration, Personalized PageRank and online LLM use. It motivates relational/associative retrieval, but this implementation does not reproduce its learned retriever or LLM components, and does not claim its reported performance.
- IRCoT (ACL 2023), official paper page: https://aclanthology.org/2023.acl-long.557/ . It interleaves retrieval and language-model reasoning for multi-step questions. Recall's explicit relation paths address a much narrower supplied-schema problem without an LLM reasoner. No IRCoT equivalence or superiority is claimed.
- LongMemEval, official repository: https://github.com/xiaowu0162/LongMemEval . It separates extraction, multi-session, temporal, updates and abstention capabilities. The repository also points to LongMemEval-V2. The included local-data loader is a retrieval-only evaluator and is not a completed benchmark run. Gold answer-session labels never enter the memory inputs.
- SQLite isolation and data-version contracts: https://www.sqlite.org/isolation.html and https://www.sqlite.org/pragma.html#pragma_data_version . They inform snapshot-consistent reads and conservative cached-view invalidation. Local regression tests additionally cover a writer committing during a reader snapshot.
What is the actual contribution here?
An engineering composition: tiny authored ranking policies; explicit topic rules; time-aware typed path execution; dependency-complete prompt packets; revalidated support certificates; and cache invalidation across edits, expiry, deletion and separate SQLite writers. None of these ingredients alone establishes a novel scientific discovery. The tests demonstrate implemented behavior under declared scope.
Evidence still required for a broader claim
Independent public benchmark results; diverse real-user paraphrases; malicious and erroneous claim ingestion evaluation; realistic multi-tenant deployment review; actual host-model answer quality; and comparisons against full existing memory systems under matched budgets. Published research results cannot be borrowed as this implementation's measurements.