Related Work¶
This page maps Memoir's design choices to the research that informs them. Sections are organized by which Memoir component the citation grounds, with a short note on relevance — not a literature review, just enough to point a curious reader at the strongest prior art.
Classifier — Hierarchical Text Classification with LLMs¶
Memoir's IntelligentClassifier performs zero/few-shot hierarchical multi-label classification with a single LLM call against a structured taxonomy prompt. This is a quietly active subfield (2024–2025) with several papers describing nearly the same construction.
- TELEClass: "Taxonomy Enrichment and LLM-Enhanced Hierarchical Text Classification with Minimal Supervision", 2024. Minimally-supervised hierarchical classification that uses LLM-generated class-indicative terms to augment the taxonomy. Closest match to Memoir's
LLMIterativeTaxonomyexpansion logic, where the taxonomy itself grows from observed content. arXiv:2403.00165 - TaxMorph: "Hierarchical Text Classification with LLM-Refined Taxonomies". Frames the LLM as a taxonomist — the model can reshape an input hierarchy into a more semantically coherent one. Justifies Memoir's stance that the taxonomy is a living artifact, not a fixed schema. arXiv:2601.18375
- Single-pass Hierarchical Text Classification with LLMs (Payberah, 2024). Single-call hierarchical classification, mirroring Memoir's "one prompt, full taxonomy in context, structured JSON out" approach. Compares flat vs. hierarchical vs. discriminative baselines. PDF
- HierPrompt: "Zero-Shot Hierarchical Text Classification", EMNLP Findings 2025. Pure zero-shot hierarchical classification — relevant because Memoir's taxonomy can be deployed cold against new domains without labelled data. PDF
- KG-HTC: "Integrating Knowledge Graphs into LLMs for Effective Zero-shot Hierarchical Text Classification", 2025. Treating the taxonomy as a structured prompt input. Same intuition Memoir uses when injecting
${TAXONOMY_BLOCK}into the classifier and Stop-hook prompts. arXiv:2505.05583 - "Effective Hierarchical Text Classification with Large Language Models", SN Computer Science, 2025. Survey-style empirical comparison of LLM-based HTC approaches. Useful background. Springer
Long-Term Memory Architecture for LLM Agents¶
The broader frame Memoir competes in: how should an agent's persistent memory be organized, retrieved, and pruned. Memoir's distinguishing claim — versionable, branchable, taxonomy-keyed memory with cryptographic proofs — sits adjacent to but not duplicated by these systems.
- Packer et al., "MemGPT: Towards LLMs as Operating Systems", 2023. The foundational hierarchical-memory-tier paper (main / recall / archival). Memoir's branches and namespaces are an alternative topology for the same problem — same goal of fitting unbounded history into a bounded working set, different solution. arXiv:2310.08560
- Xu et al., "A-MEM: Agentic Memory for LLM Agents", 2025. Current SOTA on multi-hop memory benchmarks, reporting 85–93% token-usage reduction vs. MemGPT. A-MEM's self-organized note-linking is the closest published analog to Memoir's
related_keyscross-references. arXiv:2502.12110 - Mem0: "Building Production-Ready AI Agents with Scalable Long-Term Memory", 2025. The production-systems framing Memoir competes with most directly. Useful as a comparison point on operational concerns (latency, cost, retrieval quality). arXiv:2504.19413
- "Hierarchical Memory for High-Efficiency Long-Term Reasoning in LLM Agents", July 2025. Direct support for the hierarchical-taxonomy-as-memory thesis Memoir advances.
- "Memory in the Age of AI Agents: A Survey", 2025. The survey to cite when framing Memoir's place in the landscape; identifies five mechanism families (context-resident compression, retrieval-augmented stores, reflective self-improvement, hierarchical virtual context, policy-learned management). Memoir is primarily retrieval-augmented + hierarchical virtual context with a versioning twist. Paper list
- "Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers", 2026. Recent companion survey covering evaluation methodology — useful for grounding Memoir's benchmark choices. arXiv:2603.07670
Versioned Structured Data and Content-Addressed Storage¶
Memoir's git-like-versioning-for-memory thesis sits on a body of systems work, much of it grey literature, but with one foundational academic citation.
- Benet, "IPFS — Content Addressed, Versioned, P2P File System", 2014. The canonical content-addressed-DAG paper. Memoir applies the same construction (immutable content-addressed chunks → Merkle DAG of revisions) to memory blobs rather than files. arXiv:1407.3561
Memento Pattern — Episodic, Semantic, and Spatial Memory¶
Memoir's three Memento containers (ProfileMemento, TimelineMemento, the location-keyed equivalent) map cleanly onto the cognitive-science distinction between semantic, episodic, and spatial memory that has been formalized for LLM agents over the last two years. This is the most direct theoretical alignment in Memoir.
- Park et al., "Generative Agents: Interactive Simulacra of Human Behavior", UIST 2023. Introduces the episodic-event-stream + reflection-into-semantic-memory pattern that Memoir's Timeline → Profile flow mirrors. Their three retrieval scores (recency, importance, relevance) define the design space Memoir's search engine operates in. arXiv:2304.03442
- Sumers et al., "Cognitive Architectures for Language Agents" (CoALA), 2023. The canonical academic taxonomy for agent memory; episodic / semantic / procedural categories are CoALA's. Letta, Mem0, and LangChain all use CoALA as their framing — Memoir's Mementos sit in the same ontology. arXiv:2309.02427
- "Synapse: Empowering LLM Agents with Episodic-Semantic Memory via Spreading Activation", 2025. Activation-based retrieval across the episodic ↔ semantic boundary. Closest prior art for Memoir's
related_keyscross-reference design, which is structurally a static spreading-activation graph. arXiv:2601.02744
Notable Gaps¶
Two parts of Memoir's design currently sit ahead of the published literature, worth flagging honestly:
- "Memory worthiness" filtering — the should-this-turn-become-a-memory? decision Memoir's Stop hook performs. Generative Agents and A-MEM both touch importance scoring, but no paper is dedicated to this filter.
- Multi-key sibling backlinks at storage time — A-MEM's note-linking is the closest analog, but it is computed at retrieval time, not frozen into the blob at write. Memoir's design (write-time
related_keys, edit-preserved) appears to be novel.
These are reasonable areas for a future Memoir whitepaper to claim contribution.