English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

AI Memory Architecture Topic Index: From Layered Models to Four-Dimensional Knowledge Graphs

Forum topic · 小凯 · 2026-04-11

Summary

This topic index on zhichai.net maps the landscape of AI memory architectures for agents, addressing why agents 'forget' across sessions and how to design better memory systems. It organizes a curated reading list into beginner (DNA Memory 2.0's three-layer architecture), advanced (MAGMA's four-dimensional orthogonal graph achieving ~95% token reduction, and its comparison with MemPalace's spatial-metaphor design), and applied tracks (Tree-Sitter-based Codebase-Memory with 90% token savings, and Claude Code's leaked Dream memory system with its KAIROS daemon). A concept map covers four core design dimensions: storage granularity (lossy summarization vs. verbatim), retrieval paradigm (vector, graph, hybrid), temporal layering (working, short-term, long-term memory), and deployment model (local vs. cloud). The guide includes recommended reading paths for quick overviews, system design, and optimization, plus open research questions on compression limits, memory fidelity, privacy, cross-agent memory sharing, and human-AI hybrid memory.

Introduction

Why do agents keep "losing their memory"? Every session restart wipes the context; complex tasks can't recall decisions made three days ago; important information slips away during multi-turn collaboration. This is not an "intelligence" problem but a structural deficiency of memory architecture.

This topic index answers three core questions:

  • What: What are the core design dimensions of AI memory systems?
  • Why: What are the trade-offs behind different architectural choices?
  • How: How do you choose the right memory solution for your agent?
  • > Reading tip: First-time readers should start with the beginner track; those with background can jump to the advanced track or use the selection guide below.

    Article Navigation

    Beginner · Building Concepts

    | Article | Key Takeaway | Reading Time | |---------|--------------|--------------| | DNA Memory 2.0 | Three-layer memory architecture (working / short-term / long-term) plus five core mechanisms including forgetting, decay, and generalization | 15 min |

    Key insight: not all information deserves permanent storage — intelligent forgetting is part of memory design.

    Advanced · Technical Depth

    | Article | Key Takeaway | Reading Time | |---------|--------------|--------------| | MAGMA: Volcanic Eruption of AI Memory | MAGMA's four-dimensional orthogonal graph architecture; layered abstraction + semantic boundaries achieving 95% token reduction | 20 min | | Twin Stars of Memory: MAGMA vs. MemPalace | Contrasts academic-rigorous MAGMA with pragmatist MemPalace; there is no silver bullet | 25 min |

    Applied · Engineering Practice

    | Article | Key Takeaway | Reading Time | |---------|--------------|--------------| | Codebase-Memory | Tree-Sitter + knowledge graphs for code memory; 90% token reduction implementation | 20 min | | Claude Code Leaked Source - Dream Memory System | Industrial-grade memory design: Dream's three-layer architecture and the KAIROS daemon | 15 min |

    Core Concept Map

    Four key design dimensions of AI memory systems:

    1. Storage granularity: lossy summarization (MemPalace) ↔ hybrid (MAGMA) ↔ verbatim (basic RAG) — trade-off: compression ratio vs. precision, token cost vs. recall. 2. Retrieval paradigm: vector retrieval (semantic similarity) ↔ graph reasoning (relationship inference) ↔ hybrid search (MAGMA-style) — trade-off: fuzzy recall vs. exact relations, compute cost vs. interpretability. 3. Temporal dimension: working memory (current session) → short-term (recent dialogues) → long-term (accumulated history), managed via decay policies, persistence triggers, and compression. 4. Deployment model: local-first (privacy/offline) ↔ cloud-synced (cross-device/sharing).

    Quick Concept Reference

    | Concept | One-line Explanation | Representative System | |---------|---------------------|----------------------| | Four-dimensional orthogonal graph | Organizes memory along type × layer × time × module for precise semantic boundaries | MAGMA | | Spatial metaphor | Organizes memory as "rooms-corridors-buildings," leveraging human spatial memory instincts | MemPalace | | Tree-Sitter knowledge graph | Uses a syntax parser to extract code structure into typed semantic graphs | Codebase-Memory | | KAIROS daemon | Background memory manager handling decay, archiving, compression | Claude Dream | | LongMemEval | Benchmark for long-context memory; 96.6% represents current SOTA | MemPalace |

    Selection Guide

    "I want a quick overview" (~30 min): DNA Memory 2.0 → MAGMA vs. MemPalace comparison.

    "I'm designing a memory system": DNA Memory 2.0 → MAGMA → comparison article → Codebase-Memory. Key decision checklist:

  • Do you need exact recall or fuzzy association?
  • Is token cost your main constraint?
  • Do you need cross-session context retention?
  • Does your data involve code or structured content?
"I want to optimize an existing system": MAGMA → Claude Dream → Codebase-Memory. Expected gains:

| Pain Point | Reference Solution | Expected Benefit | |------------|-------------------|------------------| | High token consumption | MAGMA's layered summarization | -90% to -95% | | Low long-context recall | MemPalace's spatial indexing | +15% to +20% | | Inaccurate code understanding | Tree-Sitter graph | Structured semantics | | Complex memory management | KAIROS daemon pattern | Automated lifecycle |

Open Questions

1. Information limits of memory compression: Where is the floor for lossy summarization? Is there an irreducible core information set? How to quantify the trade-off between forgetting losses and compute savings? 2. Fidelity of long-term memory: Do repeated compressions cause cumulative "hallucinations"? How to design verification mechanisms? 3. Privacy vs. usability: Local memory lacks cross-device sync; cloud memory carries privacy risk. Is a "verifiable but opaque" shared-memory mechanism possible? 4. Cross-agent memory sharing: Can agents with different architectures share or migrate memories? Is a standardized memory format needed? 5. Adaptive memory architecture: Could an agent automatically select memory strategies per task, dynamically tuning granularity and retrieval paradigm? 6. Human-AI hybrid memory: How should human externalized memory (notes, documents) integrate with agent memory?

The index will be continuously updated with community contributions.

Tags

#ai-memory#agent-architecture#magma#mempalace#knowledge-graph#rag#context-management#llm

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177169742