thecodacus/understory

What it solves

Understory provides a persistent, self-organizing memory layer for AI agents. It prevents agents from forgetting information across sessions and solves the problem of "knowledge rot" (orphaned or broken links) by allowing agents to maintain their own knowledge graph of facts.

How it works

Understory uses a plain-markdown system based on the Open Knowledge Format (OKF) spec. Every fact learned by an agent is stored as a markdown file with YAML frontmatter. These files are cross-linked to create a living knowledge graph.

To ensure the data remains clean, the system uses a deterministic bundle layer that enforces conformance in code rather than relying on prompts. It includes a memory_maintain tool that allows an internal agent to find and fix orphaned concepts or broken links.

Who it’s for

Developers building AI agents who need a portable, human-readable, and diffable memory system that can be integrated via MCP (Model Context Protocol) or a Web UI.

Highlights

• MCP Server: Provides tools like memory_query, memory_add, and memory_update for agents to interact with memory.
• Human-Readable: Memory is stored as plain markdown files, making it portable and compatible with git for version control.
• Visual Knowledge Graph: A web-based force-directed graph allows users to browse memory, identify orphans, and replay agent query paths.
• Local-First: Works with local models via llama.cpp or various cloud APIs (DeepSeek, OpenAI, Anthropic, Groq).
• Automatic Seed: Injects a compact overview of the knowledge base into the agent's system prompt at the start of every session.