Understory is a TypeScript-based memory system for AI agents that stores facts as plain Markdown OKF v0.1 bundles, exposing MCP, a web UI, and a CLI for querying and mutating memory. It supports local models and various LLM providers via environment variables and provides graph-based memory health tooling.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (2 so far).
What it is
understory is a memory layer for AI agents that stores learned facts as plain-Markdown concepts within an Open Knowledge Format (OKF) v0.1 bundle. It maintains a living, cross-linked knowledge graph, supports local model usage, and provides multiple access points: MCP server, a Web UI, a query-path replay, and a CLI for memory operations. The system enforces conformance in code and regenerates index files while appending log entries.
How it works
Bundles are plain Markdown with YAML frontmatter. The bundle layer ensures type conformance, regenerates index.md files, and appends log.md entries (newest-first). A deterministic process validates the bundle and sandboxes paths to the bundle root. Memory can be queried, added, updated, or maintained via memory_* tools. The web UI renders a force-directed graph and supports chat with the same agent.
Getting started
Quick start is via Docker (no clone needed). The docker-compose.yml exposes the web UI at http://localhost:3800 and MCP at /mcp, with memory mounted to /bundle. Example provider configurations illustrate using OpenAI, Anthropic, Groq, Groq-like endpoints, or local llama.cpp setups. The README provides multiple LLM configuration examples and commands to boot the service.
Recent releases
Releases: latest 0 — none
Traction
135 stars, 22 forks, 4 open issues as of creation date.
Behind the repo
No explicit startup/company link in the provided facts beyond a single project repository.
Caveats
License is not listed. Created 2026-07-09, last_push 2026-07-12. Environment notes mention AUTH_TOKEN for restricted access and GIT_AUTOCOMMIT behavior for mutations. The repository uses pnpm, Docker, and multiple environments for LLM providers.