mnemon: LLM-Supervised Persistent Memory Plugin for DeepSeek Harness
mnemon-dev/mnemon
Mnemon gives DeepSeek Harness persistent cross-session memory with graph-based intent-aware recall, in a single binary with zero API keys.
Provides cross-session graph-based memory for DeepSeek Harness, so agents never forget critical decisions.
dsh plugin --profile web add github:mnemon-dev/mnemon- Category
- Memory & Context
- Platform
- DSH-Plugin
- Author
- mnemon-dev
- Distribution
- Plugin
mnemon Key Features
mnemon Repository
mnemon Repo Summary
Mnemon is a memory and context plugin (DSH plugin) for DeepSeek Harness, maintained by mnemon-dev under the Apache-2.0 license, written in Go, and last updated in August 2026. It solves the problem of AI agents forgetting between sessions: context compaction drops critical decisions, cross-session knowledge vanishes, and long conversations push early information out of the window. Mnemon provides graph-based persistent memory with intent-aware recall, importance decay, and automatic deduplication. The entire memory path is a single local binary with zero API keys and one setup command. It works with DeepSeek Harness, Claude Code, OpenClaw, and any agent runtime, and it leverages your existing LLM subscription (e.g., Claude Max/Pro) without extra inference cost.
What core memory mechanisms does the Mnemon DSH plugin provide?
Mnemon adopts an LLM-Supervised architecture: the binary handles deterministic computation (storage, graph indexing, search, decay), while the host LLM makes judgment calls (what to remember, how to link, when to forget). It offers three primitives—remember, link, recall—forming an intent-native protocol where command names map to the LLM's cognitive vocabulary (e.g., remember not INSERT), and output is structured JSON with signal transparency. Memory is stored as four graph types (temporal, entity, semantic, causal); for example, a real knowledge graph built by Mnemon contains 87 insights and 2150 edges.
How do I install Mnemon into DeepSeek Harness?
In a DeepSeek Harness environment, install the Mnemon plugin with the following command:
dsh plugin --profile web add github:mnemon-dev/mnemon
After installation, the memory path lives at the mnemon root command, while mnemon agency ... provides Agency Preview, adding durable, project-local responsibility and effect admission to an existing Pi agent, without replacing Memory or the Agent Runtime.
How does Mnemon differ from MCP, file injection, and other approaches?
Unlike Mem0 or Letta, which embed an LLM inside the pipeline, Mnemon lets the host LLM act as an external supervisor, avoiding middlemen and extra inference cost. Compared to Claude Code Memory's file injection (reading a file at session start), Mnemon offers dynamic graph recall. Compared to MCP servers (providing tools via the MCP protocol), Mnemon fills a protocol gap for LLMs accessing databases with memory semantics; its three primitives form an intent-native protocol.
What are Mnemon's license and platform compatibility?
Mnemon is licensed under Apache-2.0, maintained by mnemon-dev, with 456 stars and 61 forks. It requires Go 1.24+, offers a Homebrew Cask installation (macOS), and is compatible with DeepSeek Harness, Claude Code, OpenClaw, and other runtimes. The project was created in February 2026 and last updated in August 2026, indicating active development.