FuRongJun-1999/dsh-memory

What it solves

It provides AI agents with long-term, cross-session memory that is reliable, auditable, and human-readable. It solves the problem of "hallucinations" and memory loss in AI agents by replacing opaque vector databases with a structured, rule-based cognitive map stored in standard Markdown files.

How it works

It functions as a standard MCP (Model Context Protocol) server. When an agent interacts with the system, it uses a set of deterministic rules to decide what to store and retrieve. Conversations are distilled into "cognitive maps" (Markdown files) that follow a specific structure (CCG six-element memory cards). The system uses a multi-layered "evidence firewall" to filter out weak or irrelevant information, ensuring that retrieved memories are verifiable and grounded in the original text rather than LLM guesswork.

Who it’s for

It is designed for developers building AI agents who need high-performance, persistent memory without the complexity of managing black-box vector databases. It is particularly useful for those using DSH, Claude Code, or any MCP-compatible agent framework.

Highlights

• Deterministic Reliability: Uses rule-based logic rather than LLM black-box decisions for memory management, ensuring results are reproducible.
• Human-Readable Storage: Memories are stored as standard Markdown files, allowing developers to manually inspect, edit, or audit the "brain" at any time.
• Framework Agnostic: Operates as a standard MCP server, making it compatible with various agent frameworks like DSH, Claude Code, and Codex CLI.
• High Performance: Features a Rust-based retrieval engine that provides near-instant memory access without third-party dependencies.
• Auditability: Every write operation requires explicit confirmation, and the system maintains logs to track what was stored, skipped, or filtered.