Keyboard shortcuts

Press ← or → to navigate between chapters

Press ? to show this help

Press Esc to hide this help

Research: Academic Foundations [SPEC]

Consolidates all academic references underpinning the Bardo Memory architecture, organized by research domain. This file is a pointer index – full annotations and competitive analysis live in the extended file.

Reader orientation: This document indexes the academic literature that grounds the Grimoire (the Golem’s persistent local knowledge base) and broader memory architecture. It belongs to the 04-memory layer. The key concept is that every design decision in the Golem’s memory system traces to published cognitive science, neuroscience, or evolutionary biology. This is a pointer index organized by research domain; annotations explain what each work argues and how it connects to the Bardo implementation. For term definitions, see prd2/shared/glossary.md.

Extended: Full specification – see ../../prd2-extended/04-memory/10-research-extended.md

Memory consolidation and forgetting

The theoretical core. Forgetting is not failure; it is regularization.

  • [EBBINGHAUS-1885] Ebbinghaus, H. Memory: A Contribution to Experimental Psychology. 1885. The forgetting curve. Half-life calibration for decay classes (episodes 48h, insights 7d, heuristics 14d, warnings 30d) derives from this.
  • [RICHARDS-FRANKLAND-2017] Richards, B.A. & Frankland, P.W. “The Persistence and Transience of Memory.” Neuron, 94(6), 2017. The paper that reframes forgetting as optimization. Memory pruning = L1 regularization. Foundational for the entire Grimoire decay architecture.
  • [WALKER-VAN-DER-HELM-2009] Walker, M.P. & van der Helm, E. “Overnight Therapy? The Role of Sleep in Emotional Brain Processing.” Psychological Bulletin, 135(5), 2009. Sleep to Forget, Sleep to Remember (SFSR) model. Grounds the Dream Engine’s REM depotentiation mechanism.
  • [STICKGOLD-2005] Stickgold, R. “Sleep-Dependent Memory Consolidation.” Nature, 437, 2005. Sleep replay strengthens memory traces selectively. Grounds NREM replay selection in dream processing.
  • [BORN-WILHELM-2012] Born, J. & Wilhelm, I. “System Consolidation of Memory During Sleep.” Psychological Research, 76, 2012. Sleep-dependent consolidation biased toward future-relevant memories. Maps to the Curator cycle’s selective promotion.
  • [NADER-2000] Nader, K. et al. “Fear Memories Require Protein Synthesis in the Amygdala for Reconsolidation after Retrieval.” Nature, 406, 2000. Reconsolidation theory: retrieved memories become labile and can be updated. Justifies the confidence-update-on-retrieval mechanism.
  • [HARDT-NADER-NADEL-2013] Hardt, O. et al. “Decay Happens: The Role of Active Forgetting in Memory.” Trends in Cognitive Sciences, 17(3), 2013. Active molecular forgetting processes. Grounds the Curator’s DOWNVOTE and pruning operations.
  • [MCCLELLAND-1995] McClelland, J.L., McNaughton, B.L., & O’Reilly, R.C. “Why There Are Complementary Learning Systems in the Hippocampus and Neocortex.” Psychological Review, 102(3), 1995. CLS theory: dual-system memory with fast hippocampal capture and slow neocortical consolidation. Grounds the Grimoire’s episodic/semantic dual-store architecture.
  • [KUMARAN-2016] Kumaran, D., Hassabis, D., & McClelland, J.L. “What Learning Systems do Intelligent Agents Need? CLS Theory Updated.” Trends in Cognitive Sciences, 20(7), 2016. Updated CLS showing replay scheduling matters: high-surprise episodes should be replayed more often. Grounds prioritized consolidation replay.
  • [OREILLY-2014] O’Reilly, R.C., Bhatt, M.A., & Russin, J.L. “Complementary Learning Systems.” Cognitive Science, 38(Suppl 1), 2014. Pattern separation (hippocampal) and pattern completion (neocortical) as complementary operations. Grounds the episodic/semantic store duality.
  • [MCCLOSKEY-COHEN-1989] McCloskey, M. & Cohen, N.J. “Catastrophic Interference in Connectionist Networks.” Psychology of Learning and Motivation, 24, 1989. The fundamental constraint that makes interleaved replay necessary. Grounds the anti-catastrophic-forgetting mechanisms.
  • [SCHAUL-2016] Schaul, T. et al. “Prioritized Experience Replay.” ICLR 2016. arXiv:1511.05952. Priority proportional to TD error magnitude. Grounds the surprise-weighted replay candidate selection in ConsolidationEngine.
  • [MATTAR-DAW-2018] Mattar, M.G. & Daw, N.D. “Prioritized Memory Access Explains Planning and Hippocampal Replay.” Nature Neuroscience, 21(11), 2018. Utility = gain * need for replay selection. Grounds the Mattar-Daw replay utility function.
  • [SELF-DEGRADATION-2025] arXiv:2505.16067. “On the Self-Degradation of Agent Memory.” 2025. Naive add-all memory consistently degrades agent performance. Incorrect past executions propagate through retrieval. Grounds the quality gate (mark_verified) on episodic consolidation.

Affect and retrieval

Emotion is not noise – it is a retrieval index.

  • [BOWER-1981] Bower, G.H. “Mood and Memory.” American Psychologist, 36(2), 1981. Mood-congruent retrieval. Directly implemented as the emotional factor (0.15 weight) in four-factor retrieval scoring.
  • [DAMASIO-1994] Damasio, A.R. Descartes’ Error: Emotion, Reason, and the Human Brain. Putnam, 1994. Somatic marker hypothesis. Implemented as the SomaticMarkerStore in the Grimoire struct. Decision-making without affect produces pathological choices.
  • [EMOTIONAL-RAG] Multiple sources. The principle that RAG systems should weight retrieved documents by emotional congruence with the query context. The Golem’s four-factor scoring is an Emotional RAG implementation.
  • [PHELPS-2004] Phelps, E.A. “Human Emotion and Memory: Interactions of the Amygdala and Hippocampal Complex.” Current Opinion in Neurobiology, 14(2), 2004. Emotion and memory as a single system with two interfaces. Grounds the emotional_tag field on every GrimoireEntry.
  • [CAHILL-MCGAUGH-1998] Cahill, L. & McGaugh, J.L. “Mechanisms of Emotional Arousal and Lasting Declarative Memory.” Trends in Neurosciences, 21(7), 1998. Arousal enhances consolidation via amygdala modulation. Grounds the arousal_encoding_factor() function.

Causal reasoning

The Golem’s causal graph is not a correlation matrix. It is a model of how the world works.

  • [PEARL-2009] Pearl, J. Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge University Press, 2009. Directed acyclic graphs for causal reasoning. The causal_edges table in SQLite is a Pearl causal graph. Interventional queries (“what would happen if I changed X?”) require causal, not correlational, knowledge.

Knowledge compression and transfer

What crosses the generational boundary, and why compression is the regularizer.

  • [SHUVAEV-2024] Shuvaev, S. et al. “Encoding Innate Ability Through a Genomic Bottleneck.” PNAS, 121(39), 2024. The genome is ~1000x smaller than the information needed for brain connectivity, yet organisms have innate behaviors. Compression IS the regularizer. Grounds the 2048-entry genomic bottleneck in death bundles.
  • [HEARD-MARTIENSSEN-2014] Heard, E. & Martienssen, R.A. “Transgenerational Epigenetic Inheritance: Myths and Mechanisms.” Cell, 157(1), 2014. Epigenetic marks fade within 2-3 generations. Grounds the 0.85^N generational confidence decay.
  • [HINTON-NOWLAN-1987] Hinton, G.E. & Nowlan, S.J. “How Learning Can Guide Evolution.” Complex Systems, 1, 1987. The Baldwin Effect: what transfers is not knowledge but the capacity to learn. Grounds the PLAYBOOK.md inheritance model.
  • [WEISMANN-1893] Weismann, A. The Germ-Plasm: A Theory of Heredity. 1893. Somatic/germline barrier. Grounds the architectural separation between what dies with the Golem and what crosses to successors.

Testing effect and spacing

How retrieval itself is the learning mechanism.

  • [ROEDIGER-KARPICKE-2006] Roediger, H.L. & Karpicke, J.D. “Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention.” Psychological Science, 17(3), 2006. Retrieval strengthens memory traces more than re-study. Grounds the Curator’s re-validation requirement and the strength-increment-on-positive-outcome mechanism.
  • [CEPEDA-2006] Cepeda, N.J. et al. “Distributed Practice in Verbal Recall Tasks.” Psychological Bulletin, 132(3), 2006. Spaced retrieval > massed retrieval. Grounds the 50-tick Curator cycle interval.

Agent knowledge systems

The ML/AI systems that the Grimoire draws from and extends.

  • [REFLEXION-2023] Shinn, N. et al. “Reflexion: Language Agents with Verbal Reinforcement Learning.” NeurIPS 2023. Self-critique as episodic memory. Grounds the Reflexion pipeline.
  • [EXPEL-2023] Zhao, A. et al. “ExpeL: LLM Agents Are Experiential Learners.” AAAI 2024. Cross-episode pattern extraction. Grounds the ExpeL distillation stage.
  • [VOYAGER-2023] Wang, G. et al. “Voyager: An Open-Ended Embodied Agent with Large Language Models.” arXiv:2305.16291, 2023. Store knowledge as text injectable into LLM prompts. Grounds the content: String field and the Skill Sandbox ingestion stage.
  • [COALA-2023] Sumers, T. et al. “Cognitive Architectures for Language Agents.” TMLR, 2024. Three-store memory split (episodic/semantic/procedural). Grounds the LanceDB/SQLite/PLAYBOOK.md architecture.
  • [A-MEM-2024] Xu, Z. et al. “A-MEM: Agentic Memory for LLM Agents.” arXiv:2502.12110, 2025. Bi-temporal metadata for time-aware retrieval. Grounds the valid_from/valid_until fields.
  • [MEM0-2025] Chhikara et al. arXiv:2504.19413, 2025. Two-phase extraction-update pipeline achieving 26% higher accuracy, 91% lower p95 latency, 90% token savings vs. OpenAI memory. Validates the tiered local Grimoire architecture.
  • [GRAPHITI-2025] Rasmussen et al. arXiv:2501.13956, 2025. Bi-temporal knowledge graph with episode/semantic/community subgraphs. 94.8% accuracy on DMR vs. MemGPT 93.4%. Grounds the causal_edges temporal graph in SQLite.
  • [COALA-EPISODIC-2025] “Episodic Memory is the Missing Piece for Long-Term LLM Agents.” arXiv:2502.06975, 2025. Episodic memory enables single-shot learning from unique events — critical in DeFi where market events don’t repeat.
  • [MEMACT-2025] arXiv:2510.12635, 2025. Learned Prune&Write operator enables 14B model to match 235B model accuracy using 49% average context. Grounds the Curator’s DOWNVOTE/pruning operations.
  • [ACE-2025] Zhang et al. “ACE: Agentic Context Engineering.” arXiv:2510.04618, 2025. Generator–Reflector–Curator cycle for context evolution. +10.6% on AppWorld. Directly maps to the Grimoire’s three-loop learning architecture.
  • [ARIGRAPH-2024] Anokhin et al. arXiv:2407.04363, 2024. Semantic + episodic memory as knowledge graph world model. Outperforms established memory methods and RL baselines. Grounds the causal graph architecture.

Retrieval strategies

How the right information reaches the context window at the right time.

  • [SELF-RAG-2023] Asai, A. et al. “Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection.” arXiv:2310.11511, 2023. Adaptive retrieval decisions via reflection tokens. Grounds the should_retrieve() heuristic that skips retrieval when context is sufficient.
  • [RAPTOR-2024] Sarthi, P. et al. “RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval.” ICLR 2024. arXiv:2401.18059. Hierarchical summarization tree for multi-scale retrieval. Grounds the episodic/cluster/semantic three-level retrieval hierarchy.
  • [COLBERT-2020] Khattab, O. & Zaharia, M. “ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT.” SIGIR 2020. arXiv:2004.12832. Per-token embeddings with MaxSim scoring. Grounds the ColBERT-style reranking step that catches term-specific matches.
  • [PARK-2023] Park, J.S. et al. “Generative Agents: Interactive Simulacra of Human Behavior.” UIST 2023. arXiv:2304.03442. Three-factor retrieval formula (recency, importance, relevance). Ablation shows removing any single factor causes behavioral degeneration. Grounds the four-factor scoring.
  • [RRF-2009] Cormack, G.V., Clarke, C.L.A., & Butt, S. “Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods.” SIGIR 2009, 758-759. Grounds the RRF score fusion across vector, keyword, and graph retrieval backends.
  • [STREAMINGRAG-2024] Arefeen, M.A. et al. “StreamingRAG: Real-time Contextual Retrieval-Augmented Generation.” ACM SIGMOD AI4Sys Workshop, 2024. 5-6x throughput improvement via incremental graph extension. Grounds streaming index updates at Gamma rate.
  • [IRAG-2024] Arefeen, M.A. et al. “iRAG: Incremental Retrieval-Augmented Generation for Streaming Data.” arXiv:2404.12309, 2024. 23-25x faster ingestion via deferred extraction. Grounds the lazy episode enrichment pattern (extract detailed context only when retrieved).
  • [MEMORYBANK-2024] Zhong, W. et al. “MemoryBank: Enhancing Large Language Models with Long-Term Memory.” AAAI 2024. Long-term memory augmentation for LLMs. Grounds persistent knowledge store design.

Temporal knowledge graphs

Relational, temporal, structured memory for DeFi topology.

  • [SNODGRASS-AHN-1985] Snodgrass, R.T. & Ahn, I. “A Taxonomy of Time in Databases.” Proceedings ACM SIGMOD, 1985, 236-246. Foundational bi-temporal database theory. Two timestamps needed: valid time (when true) vs. transaction time (when learned). Grounds the TemporalTriple’s valid_from/valid_to/recorded_at fields.
  • [JENSEN-SNODGRASS-1999] Jensen, C.S. & Snodgrass, R.T. “Temporal Data Management.” IEEE Trans. Knowledge and Data Engineering, 11(1), 1999. Extended bi-temporal formalization. Grounds temporal range queries.
  • [KITZLER-2022] Kitzler, S. et al. “Disentangling DeFi Compositions.” ACM Transactions on the Web, 16(4), 2022. DEX and lending protocols have highest centrality in composition graph. Multi-hop dependency chains invisible in single-transaction analysis. Grounds the DeFi topology model and k_hop_neighbors traversal.
  • [GRAPHRAG-2024] Edge, D. et al. “From Local to Global: A Graph RAG Approach to Query-Focused Summarization.” arXiv:2404.16130, 2024. Community-level summarization from knowledge graphs. Grounds the graph-based retrieval backend.
  • [DEXPOSURE-2024] “A Large-Scale Dataset for Inter-Protocol Credit Exposure in DeFi.” arXiv:2511.22314, 2024. Inter-protocol credit exposure data. Supports the heterogeneous graph model for DeFi topology.

Context compression and retrieval quality

How context engineering interacts with the memory system.

  • [ACON-2025] Kang et al. “ACON: Adaptive Compression.” arXiv:2510.00615, 2025. Failure-driven compression guideline optimization: 26–54% peak token reduction. The Curator’s context injection learns from retrieval failures.
  • [LLMLINGUA-2023] Jiang et al. “LLMLingua: Compressing Prompts for Accelerated Inference.” EMNLP 2023. Up to 20x compression via small-LM importance scoring. Applicable to episode summaries. Grounds the PromptCompressor implementation.
  • [LONGLLMLINGUA-2024] Jiang, H. et al. “LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression.” ACL 2024. arXiv:2310.06839. Question-aware compression achieving 21.4% performance improvement with 4x compression. Grounds the question_aware_compress() function.
  • [LOST-IN-MIDDLE-2023] Liu, N.F. et al. “Lost in the Middle: How Language Models Use Long Contexts.” TACL 2024. arXiv:2307.03172. U-shaped attention curve: LLMs best attend to beginning and end of context. Grounds the reorder_for_u_curve() block placement strategy.
  • [MEMGPT-2023] Packer, C. et al. “MemGPT: Towards LLMs as Operating Systems.” arXiv:2310.08560, 2023. Structured memory blocks with clear labels improve LLM reasoning by 30-60%. Grounds the ContextBlock/ContextAssembler typed block architecture.
  • [ATTENTION-SINKS-2024] Xiao, G. et al. “Efficient Streaming Language Models with Attention Sinks.” ICLR 2024. arXiv:2309.17453. Attention sink tokens for streaming contexts. Relevant to I-frame/P-frame delta compression.
  • [GIST-TOKENS-2023] Mu, J. et al. “Learning to Compress Prompts with Gist Tokens.” NeurIPS 2023. Learned compression tokens for prompt compression. Complementary to perplexity-based compression.
  • [TABULAR-FORMATTING-2024] arXiv:2412.17189. “Tabular Formatting Improves LLM Performance on Data-Analytics Tasks.” 2024. 40.29% average performance gain with tabular formatting. Grounds the recommendation to format pool data as markdown tables.
  • [OBSERVATION-MASKING-2025] Lindenbauer et al. arXiv:2508.21433, NeurIPS 2025 DL4C Workshop. Simple masking of all but M most recent observations halves cost while matching LLM summarization. The Curator’s triage (“preserve/abstract/forget”) is more effective than LLM summary.
  • [WEIGHTED-RETRIEVAL-2025] “Weighted Memory Retrieval.” Frontiers in Psychology, 2025. Trained ACAN replaces static retrieval weights. Points toward eventual learning of the four-factor scoring weights.
  • [CAUSALPROBE-2024] Chi et al. NeurIPS 2024. LLMs perform only level-1 (associative) causal reasoning by default. G²-Reasoner with RAG pushes toward level-2 inference. Grounds why the Grimoire stores causal graphs rather than relying on LLM causal reasoning alone.

Forgetting as defense and knowledge poisoning

The defense layer research foundation.

  • [POISONED-RAG-2024] Zou et al. “PoisonedRAG.” 2024. Five malicious texts achieve 90% attack success against naive RAG. Grounds the eight-layer defense stack in 10-safety.md.
  • [COGNITIVE-BACKDOORS-2024] Multiple sources. Backdoor injection via vector stores. Grounds the QUARANTINE stage in four-stage ingestion.
  • [RAG-DEFENDER-2025] Multiple sources. Outlier isolation for retrieval results. Grounds the Bloom Oracle and immune memory architecture.