Research And Evaluation
CoreTex uses reproducible memory evaluation to support continuous improvement. Its purpose is to expose enough safe, portable surface area for repeated miner contributions.
Research foundations
The design draws on several established lines of work. CoreTex selects ideas from them while retaining its own architecture.
| Research | Relevance to CoreTex |
|---|---|
| Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks | Separating non-parametric evidence retrieval from generation, with provenance available to the reader |
| Lost in the Middle: How Language Models Use Long Contexts | Why selecting and packing a small amount of useful context can be preferable to supplying an undifferentiated long history |
| Generative Agents: Interactive Simulacra of Human Behavior | Memory streams, relevance/recency/importance retrieval, and periodic reflection over accumulated experience |
| MemGPT: Towards LLMs as Operating Systems | Treating memory management and movement between persistent state and the active context as an explicit system layer |
| MemoryBank: Enhancing Large Language Models with Long-Term Memory | Continuous memory updates, salience, reinforcement, and selective retention over sustained interaction |
| HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models | Entity/relation organization and graph-supported multi-hop retrieval alongside similarity search |
| Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory | A practical extraction, consolidation, storage, and retrieval pipeline, including explicit accuracy/latency/token-cost trade-offs |
| Hindsight is 20/20: Building Agent Memory that Retains, Recalls, and Reflects | Separating retained evidence, experiences, summaries, and evolving beliefs while keeping reflection available as an operational phase |
These ideas appear in CoreTex as separate write and read stages, explicit M3 consolidation, structured relations, temporal validity, provenance, final context budgeting, and an adapter that remains outside the model itself.
Evaluation references
The public benchmark programme emphasizes abilities that a useful agent memory must retain as its history grows and changes:
| Benchmark | What it tests |
|---|---|
| LongMemEval | Information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention |
| LoCoMo | Very long-term conversational memory, temporal and causal reasoning, and event summarization |
| MEME | Evolving multi-entity state, dependency cascades, absence, and deletion |
| STALE | Implicit invalidation, current-state resolution, false-premise resistance, and policy adaptation |
| LongMemEval-V2 | Memory for environment state, workflows, recurring failure modes, premise awareness, and the accuracy/latency trade-off |
The runtime and neutral adapter have been exercised on temporal updates, conflicts, paraphrases, multi-hop relations, consolidation, tool/document memory, packing, deletion, validity, and restart behavior. Neutral interfaces also support comparisons with external memory systems on public benchmarks such as LongMemEval, MEME, and STALE when their native dependencies and evaluation conditions are pinned.
Preliminary external probes provide diagnostic evidence with sample sizes too small for leaderboard ordering. They have shown competitive retrieval and clear remaining runway, especially in update-aware consolidation, ranking, and efficient packing. Some external systems obtain higher raw recall by returning stale or otherwise invalid evidence. CoreTex reports validity-gated results separately and excludes such evidence from useful memory.
The benchmark uses a common ordered history, query, context budget, and reader interface for every system through native adapters. CoreTex-specific checks cover protocol properties required for mining, including deterministic replay, portability, state continuity, and the fixed safety substrate.
The intended long-term process is incremental: miners improve M1-M6 behavior, accepted work advances the shared state, validators reproduce that decision, and portable installations synchronize to the new release while their application integration remains stable.