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

Introduction: Why Do Agents Keep "Forgetting"?

When a conversation restarts, prior context vanishes; on complex tasks, the agent cannot recall decisions made three days ago; across multi-turn collaboration, important information slips away. This is not an "intelligence" problem but a structural flaw in memory architecture.

This topic series answers three core questions:

• What: What are the core design dimensions of AI memory systems?
• Why: What trade-offs drive different architectural choices?
• How: How do you pick the right memory solution for your Agent?

Article Navigation

Beginner · Building Concepts

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

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 achieve ~95% token reduction | 20 min | | Twin Stars of Memory: MAGMA vs. MemPalace | Design philosophy contrast between the academically rigorous MAGMA and the pragmatic MemPalace; no silver bullets | 25 min |

Applied · Engineering Practice

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

Core Concept Map

• Storage granularity: lossy summarization (MemPalace) ↔ hybrid (MAGMA) ↔ verbatim (RAG baseline) — trade-off: compression vs. precision, token cost vs. recall.
• Retrieval paradigm: vector search (semantic similarity) ↔ graph reasoning (relational inference) ↔ hybrid search (MAGMA-style).
• Temporal dimension: working memory (current session) → short-term (recent dialogues) → long-term (accumulated history), with lifecycle management (decay, persistence triggers, compression).
• Deployment: local-first (privacy/offline) ↔ cloud-synced (cross-device/sharing).

Key Concepts Quick Reference

| Concept | One-line Explanation | Representative | |---|---|---| | Four-dimensional orthogonal graph | Organize memory by type × hierarchy × time × module for precise semantic boundaries | MAGMA | | Spatial metaphor | Memory as "rooms-corridors-buildings", leveraging human spatial memory | MemPalace | | Tree-Sitter knowledge graph | Parse code structure into a typed semantic graph | Codebase-Memory | | KAIROS daemon | Background process handling decay, archiving, compression | Claude Dream | | LongMemEval | Long-context memory benchmark; 96.6% represents current SOTA | MemPalace |

Reading Path Guide

• Quick overview (~30 min): DNA Memory 2.0 → MAGMA vs. MemPalace comparison.
• Designing a memory system: DNA Memory 2.0 → MAGMA → comparison → Codebase-Memory. Decision checklist: exact recall vs. fuzzy association? Token cost a constraint? Cross-session context needed? Code/structured data involved?
• Optimizing an existing system: MAGMA (95% token reduction) → Claude Dream (industrial architecture) → Codebase-Memory (domain-specific techniques).

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

Open Questions

1. Where is the information boundary of memory compression? Is there an incompressible core information set? How to quantify the trade-off between forgetting loss and compute savings? 2. How to guarantee long-term memory "truthfulness"? Can repeated compression accumulate hallucinations? What verification mechanisms ensure accuracy? 3. Privacy vs. usability: Can we design "verifiable but opaque" shared memory? 4. Cross-agent memory sharing: Can agents with different architectures share or migrate memories? Is a standardized memory format needed? 5. Adaptive memory architecture: Can agents automatically select memory strategies (granularity, retrieval paradigm) by task type? 6. Human-AI hybrid memory: How should agent memory integrate with human externalized memory (notes, documents)?

The index is continuously updated; community contributions are welcome.