Research Frame

CoreTex reflects several recurring findings from retrieval and agent-memory research:

Theme	Design response
Dense embeddings miss memory-specific relevance	Qwen reranking and graded qrels check answer-bearing evidence beyond surface similarity
Long-horizon memory changes over time	Temporal validity, supersession, and stale-memory rejection are first-class substrate surfaces
Multi-hop agent memory needs structure	Relation and category-routing regions make useful paths explicit
Retrieval quality depends on hard negatives	Corpus generation includes plausible wrong documents and near-collision cases
Memory systems need compression	The 32 KB substrate forces miners to encode useful routing behavior under a fixed state budget
Public benchmarks invite overfitting	Hidden packs, canary records, seed reveal, and post-reveal replay separate search from verification

Representative references:

Reference	Relevance to CoreTex
MemReranker: Reasoning-Aware Reranking for Agent Memory Retrieval	Memory retrieval needs calibrated relevance, temporal reasoning, causal reasoning, and coreference handling. CoreTex reflects those families in qrels, hidden packs, and policy atoms.
Qwen3-Reranker-0.6B	Public launch reranker used for query/document ranking, pinned by model id, revision, prompt hash, and runtime.
On the Theoretical Limitations of Embedding-Based Retrieval	Single-vector retrieval has structural limits under realistic relevance sets. CoreTex uses compact routing state, relation surfaces, temporal atoms, and reranking.
WARP: An Efficient Engine for Multi-Vector Retrieval	Multi-vector and late-interaction retrieval motivate richer routing signals than one document embedding.
LoCoMo: Evaluating Very Long-Term Conversational Memory of LLM Agents	Long-term dialogue evaluation motivates current/stale facts, temporal grounding, and long-range consistency checks.
MemoryAgentBench	Incremental multi-turn memory benchmarks emphasize accurate retrieval, update behavior, long-range understanding, and selective forgetting.
MemoryArena	Multi-session agent tasks connect memory quality to later decisions, which supports evaluating memory as useful routing state.
Experience Compression Spectrum	Agent memory can be treated as compressed experience. CoreTex makes compression explicit through a fixed 32 KB substrate.
BEIR	Retrieval evaluation context for heterogeneous tasks and metrics such as nDCG, recall, and MRR.
MTEB	Broader embedding and retrieval benchmark context for model comparison and reproducible evaluation.

CoreTex turns these concerns into protocol mechanics: compact state, public roots, hidden evaluation, pinned scoring, epoch reveal, and independent replay.