Memsmith

MemSmith

Persistent, shared memory for AI coding agents. Built for Claude Code.

What it does • Install • Quick start • Capabilities • Runtimes • Troubleshooting • License

MemSmith gives your AI coding agent memory that survives between sessions. It records what the agent does, turns that activity into short typed observations (decisions, bug fixes, discoveries, and changes), and lets the agent search them in later sessions. In team mode, a whole team shares one searchable memory.

What MemSmith does

1. Capture. Claude Code hooks send prompts, tool use, file edits, and session summaries to a MemSmith server. MemSmith redacts known secrets before it stores anything.
2. Generate. A provider that you choose (Claude, Gemini, OpenRouter, or Ollama) turns the activity into observations. If a hosted provider fails, MemSmith can run the job on a local Ollama model.
3. Store. Observations go into Postgres with pgvector. A local embedder produces the vectors, so semantic search needs no API key.
4. Recall. Claude searches memory with MCP tools before it searches the code or the web. You can also search from a terminal or the dashboard.

Install

Requires Node.js 20.12 or later. MemSmith installs Bun and uv if they're missing. You don't need Docker.

npx memsmith install

Or install from the plugin marketplace inside Claude Code:

/plugin marketplace add shshalom/memsmith
/plugin install memsmith@shshalom

The installer asks which generation provider to use. The default, Claude Haiku 5.5, needs an Anthropic API key. To use Ollama only, choose it at the prompt. For all options, see Installation.

To update, run npx memsmith@latest update.

Quick start

1. Restart Claude Code, and work as usual. MemSmith captures your sessions in the background. Context from earlier sessions appears from your second session in a project.

2. Ask Claude about past work, such as "why did we choose this approach?". Claude searches memory with observation_search and observation_get.

3. Search from a terminal, in a project folder:

   memsmith search "start at login"
   memsmith get <id>
   

4. Save a note on purpose:

   memsmith note "Deploys freeze on Fridays"
   

5. Open the dashboard at the address that memsmith status prints.

6. If something looks wrong, run memsmith doctor. Add --fix to repair what it finds.

Without a global install (npm install -g memsmith), prefix each command with npx.

Capabilities

Runtimes

MemSmith runs the same engine in two runtimes. Runtimes compares them.

A project moves from local to server through the GO TEAM flow in the dashboard. The flow copies local data to the team server, verifies the copy, and then switches the project. MemSmith never deletes local data. A teammate who clones the repository joins the team from the dashboard with a team key. For deployment, see Team server and the AWS runbook.

Troubleshooting

Run memsmith doctor first. For fixes to common problems, such as a stopped server, missing context, or a blocked tool call, see Troubleshooting.

Documentation

The documentation source is in docs/public/. Start with the introduction.

Development

To build from source, clone the repository, run npm install, and then run npm run build. To run the tests, run bun test tests/server/ for a fast signal. After a source change that lands in a plugin bundle, run npm run build and commit the regenerated files under plugin/. For contributor details, see Development.

Report issues at GitHub Issues. Security reports go through GitHub private vulnerability reporting; see SECURITY.md.

License

MemSmith is licensed under the Apache License 2.0. See the LICENSE file for the full text, and docs/license.md and docs/ip-boundary.md for the licensing scope.

Credits

MemSmith includes software derived from an Apache-2.0 licensed open source project. See the NOTICE file for attribution.