Research And Framing
CoreTex follows a specific line of recent retrieval and agent-memory research. The design is not a direct implementation of any one paper; it is a protocol design that takes the recurring failure modes seriously: fixed-vector retrieval limits, stale memory, hard negatives, multi-hop relation structure, and long-horizon compression pressure.
| Work | What it contributed to CoreTex |
|---|---|
| MemReranker: Reasoning-Aware Reranking for Agent Memory Retrieval | Identifies why generic semantic rerankers fail in memory settings: semantically similar fragments may not contain answer-bearing information. The paper's 0.6B/4B reranker family, temporal/causal/coreference focus, and hard-sample discrimination directly match CoreTex's benchmark families. |
| Qwen3-Reranker-0.6B | Provides the practical launch reranker: a public 0.6B text-ranking model with long-context support and CPU-deployable Hugging Face weights. |
| On the Theoretical Limitations of Embedding-Based Retrieval | Grounds the decision not to rely on one dense vector. CoreTex uses many compact retrieval keys, relation routes, temporal metadata, and reranking because single-vector retrieval has structural limits under growing corpora. |
| WARP: An Efficient Engine for Multi-Vector Retrieval | Supports the multi-vector / late-interaction direction: retrieval quality can depend on many small interaction points rather than a single global embedding. |
| LoCoMo: Evaluating Very Long-Term Conversational Memory of LLM Agents | Motivates temporal current/stale evaluation. Remembering the old fact is not enough; a useful memory system must retrieve the current fact and suppress superseded information. |
| MemoryAgentBench | Reinforces incremental multi-turn memory as an evaluation target: the correct memory state changes over time and must be updated rather than merely appended. |
| MemoryArena | Frames memory as something agents use to guide later decisions across interdependent sessions, not as isolated recall. This supports CoreTex's long-horizon compression pressure. |
| Experience Compression Spectrum | Positions memory, skills, and rules as different compression levels. CoreTex's 32 KB substrate is a protocol-level compression surface: miners compete over what should survive into the shared state. |
| MTEB and BEIR | Provide the broader information-retrieval context for metrics such as nDCG, MRR, recall, and heterogeneous retrieval evaluation. |
The consistent theme is that more context is not the same as better memory. Good memory requires retrieval structure, temporal discipline, relation structure, and compression. CoreTex makes those properties measurable and economically legible.