What it does
memdex stores file summaries keyed to a hash of each file's content, so a summary is reused until the file changes and the agent is told when it goes stale. Notes on decisions, pitfalls and handoffs cite the files they rely on and are flagged when those files change. Agents can claim the files they are editing, post messages to other agents, read their inbox and leave handoffs, and all worktrees of a repository share one memory. The publisher's benchmark on a 66-file repository estimates 154,105 tokens to read every file against 3,471 tokens for file outlines, using a rough characters-divided-by-four count.
It suits several agents or sessions working on the same repository, where each new session would otherwise re-read the same files and rediscover the same pitfalls. The publisher names Claude Code, Codex, Cursor and DeepSeek Harness among compatible clients. Any MCP client can connect, and a model that can only run shell commands can join without configuration.
The connect command detects supported clients that are already installed and merges the memdex entry into their config files, saving a backup first. It can list, dry-run or remove the entry. Its serve-check option starts the server from the written entry and confirms it answers MCP initialize and tools/list; it does not confirm that the client loaded the server. Each agent is given its own MEMDEX_AGENT name. For team use, the config can be committed to the repository, but each teammate needs memdex available on their PATH.
Limitations and requirements
Requires the memdex Python package to be installed and available on PATH. The claude CLI is optional; without it, Claude Code configuration is printed rather than run. Harness setup with more than one profile needs an explicit profile name. Malformed JSON or TOML configs are left unchanged, and a Zed settings file containing comments is not rewritten but gets the block printed for manual pasting. Token savings are estimates.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.ChoYugyeong/memdex
Shared memory for AI coding agents: summaries, symbol map, notes and agent-to-agent messages.
v0.3.1As shown on registry.modelcontextprotocol.io when we last read it