MCP Smart Context: Building a Human-like Cognitive Memory Architecture for AI

The MCP Smart Context project proposes an AI cognitive memory architecture based on a three-layer memory hierarchy, aiming to address the limitations of traditional naive RAG methods in complex workflows and multi-turn dialogues. This architecture simulates human cognitive mechanisms, enabling persistent workspaces, intelligent token budget management, and autonomous knowledge base maintenance, and drives AI from passive information retrieval to active cognitive management.

With the rapid development of LLMs today, traditional naive RAG methods (fragmented document splitting + vector search) are inadequate for complex development workflows and multi-turn dialogues. The core issue is their inability to effectively manage context information, leading to wasted cognitive resources and loss of key information, which limits the practicality of AI agents.

First Layer: Working Memory (L1 Working Memory)

• Strict token budget control: Enforce maximum token limits based on IDE configurations
• Semantic compression mechanism: Automatically strip implementation details while retaining function signatures and interfaces when approaching the budget
• Context snapshots: Support saving/restoring current mental states to enable task switching

Second Layer: Short-Term Memory Management (L2 Short-Term Memory)

• Heuristic scoring system: Calculate eviction scores based on semantic distance and time decay
• Human-machine collaboration confirmation: Prompt users for confirmation when evicting weakly relevant contexts

Third Layer: Long-Term Memory (L3 Long-Term Memory)

• AI knowledge wiki: Autonomously generate and maintain an Obsidian-style knowledge base
• Dual-layer AST indexer: Build lightweight symbol mappings using tree-sitter

Perception and Discovery Tools

• index_workspace: Scan the root directory to index AST structures
• search_in_files: Search interface supporting Glob and regular expressions
• read_ast_index: Scan architecture without loading full files
• search_wiki/read_article: Retrieve from long-term memory storage

Attention and Context Management Tools

• view_file/read_chunk: Pull specific information into working memory
• pin_context/unpin_context: Manage key file persistence
• drop_context: Manually clear memory files

Metacognition and Workspace Control Tools

• compact_context: Force summarization of inactive files to free up tokens
• plan_eviction: Consult the eviction engine to find stale contexts
• snapshot_context/restore_context_snapshot/list_snapshots: Save and load workspace states

• Autonomous wiki management: AI can execute write_article/update_links to maintain the knowledge base
• Precise chunk loading: Load specific line ranges via read_chunk
• Stateful workspace: Support instant switching between multiple debugging sessions
• AST-driven discovery: Fast and memory-efficient symbol parsing
• Security protection: Built-in path traversal and Shell injection protection

• Supported IDEs: Antigravity/Gemini, Claude Code, Cursor, VS Code, Windsurf
• Installation and configuration: Built-in interactive wizard that prompts for parameters like OpenAI API key (optional) and token budget
• Notes: File monitoring is disabled by default to prevent CPU spikes; can be manually enabled for pure CLI workflows

MCP Smart Context marks a shift in AI context management from passive retrieval to active cognition. AI agents can build deep project understanding and maintain continuity, evolving from one-time Q&A machines to intelligent collaborators that accumulate knowledge. It brings qualitative improvements to long-term maintenance of complex projects and multi-round iterative tasks, making AI a partner with memory and accumulation.

MCP Smart Context provides AI with a context management solution close to human cognition through its three-layer memory hierarchy. From fine control of working memory to autonomous maintenance of long-term knowledge, it demonstrates a new direction in AI engineering. For developers pursuing deep AI collaboration, this project is not only a technical tool but also a philosophical practice on AI memory and learning, worthy of in-depth research and trial.