memry

memry gives your AI coding agents a persistent memory that carries over between sessions and projects. Decisions, bug fixes and discoveries saved in one session are available in the next one, so you do not have to explain the same context again.

Requirements

• macOS, Linux or Windows (amd64 or arm64)

• At least one supported agent. For Claude Code, its CLI (claude) must be on your PATH.

Related MCP server: Muninn

Install

With Homebrew:

brew install mrtheroi/tap/memry
memry setup

Other ways to install memry, a single binary with no other dependency:

• With Go: go install github.com/mrtheroi/memry-cli/cmd/memry@latest. The binary lands in $(go env GOPATH)/bin, which must be on your PATH.

• From a release: download the memry_<version>_<os>_<arch>.tar.gz archive for your system, check it against checksums.txt, and put the memry binary from it on your PATH, for example:

  tar -xzf memry_1.0.0_darwin_arm64.tar.gz memry
  sudo install -m 0755 memry /usr/local/bin/memry

On Windows, see Windows.

Your agents run memry from the path it is installed at, so outside Homebrew keep the binary where it is, or run memry setup again after you move it.

memry setup asks for your email, sends you a 6-digit login code, then asks which agents you use and connects memry to each one you select (the ones found on your machine are preselected; press space to select, enter to confirm). Start a new session in your agent afterwards to use it.

To choose the agents without the prompt, for example in a script, pass their keys:

memry setup --agents=claude-code,codex

Run memry setup again to change the selection; memry is removed from agents you deselect.

Claude Code plugin

This repository is also a Claude Code plugin marketplace. Its memry plugin adds a /memry:setup skill that walks you through installing the CLI and running memry setup.

From your terminal:

claude plugin marketplace add mrtheroi/memry-cli
claude plugin install memry@memry

Or inside a Claude Code session:

/plugin marketplace add mrtheroi/memry-cli
/plugin install memry@memry

On Claude Code 2.1.275 or later, one in-session command does both: /plugin install memry --marketplace mrtheroi/memry-cli.

The plugin does not register the MCP server itself; memry setup does, so memry is never connected twice.

Self-hosted

On a self-hosted memry server, your admin creates a token for you instead of the email login. Connect with:

memry setup --url https://memry.company.internal --token

--url is required with --token, so the token is never sent to the public server. Setup asks for the token without showing it, checks it with the server before saving it, then asks which agents you use as usual. For scripts, set the MEMRY_TOKEN environment variable: with --token and no value, setup uses it instead of the prompt, also with --no-interaction, and the token stays out of the process list. --token=<token> also skips the prompt (and wins over MEMRY_TOKEN), but the token then lands in your shell history and the process list.

MEMRY_TOKEN="$token" memry setup --url https://memry.company.internal --token --no-interaction

MEMRY_TOKEN is only read together with --token, so it never changes the email login.

Use an https:// URL. Setup sends your token or login code to the server, and memry mcp sends the saved token with every request, so a plain http:// URL is only accepted for this machine (localhost, a 127.x.x.x address or [::1]). For any other http:// server setup stops before asking or sending anything: "Refusing to send credentials over plain HTTP to . Use an https:// URL, or pass --insecure to allow it (for example on a trusted LAN).". Pass --insecure only on a network you trust: the credentials then travel in clear text. The same rule applies to MEMRY_URL.

Windows

Install with WinGet, then run setup in a new terminal so the memry command is on your PATH:

winget install Memry.Memry
memry setup

Other ways to install it:

• With Go: go install github.com/mrtheroi/memry-cli/cmd/memry@latest. The binary lands in %GOPATH%\bin (go env GOPATH), which must be on your PATH.

• From a release: download memry_<version>_windows_<arch>.zip (amd64 or arm64), check it against checksums.txt (signed with cosign, see the cosign command in the release process), and put memry.exe from it on your PATH.

Update with winget upgrade Memry.Memry. To uninstall, run memry uninstall, then winget uninstall memry; for a manual install, memry uninstall prints the path of the memry.exe to delete.

What differs from macOS and Linux:

• Config file. %USERPROFILE%\.config\memry\config.json, readable only by your user. memry finds your home in USERPROFILE, then HOME, then two levels above LOCALAPPDATA; without any, it stops with "Could not find your home directory; set USERPROFILE or HOME." (on every OS, an unset home is an error). Use --config or MEMRY_CONFIG to put it elsewhere.

• Claude Code. Setup registers memry as a local memry mcp server (claude mcp add --transport stdio --scope user memry -- <memry.exe> mcp) and adds the same SessionStart hook. Both claude.exe and claude.cmd are supported.

• The hook runs memry by its name, so memry must be on your PATH. WinGet guarantees it; with go install, add %GOPATH%\bin to your PATH. Claude Code hands its own PATH to the hook, so after installing, restart the terminal or editor (VS Code, JetBrains) you start Claude Code from; a window opened before the install does not see memry, and the hook then loads no context without any error. If you run Claude Code with CMD instead of Git Bash, CMD searches the current directory first, so do not start sessions in a folder that contains an untrusted memry.exe.

• Antivirus and SmartScreen. memry.exe is not Authenticode-signed, so SmartScreen or your antivirus may warn about it. To check a download, verify the signed checksums.txt as described in CONTRIBUTING.md and compare the archive's SHA-256.

• Terminals without color support. In a legacy console, the agent question is replaced by Agents: <list> (pass --agents to choose). and keeps the defaults; use --agents to choose.

Custom config file

By default memry keeps your login in ~/.config/memry/config.json. To use another file, set the MEMRY_CONFIG environment variable or pass the global --config <path> (or --config=<path>) option to setup, uninstall and delete-account:

memry setup --config ~/work/memry.json

The precedence is --config, then MEMRY_CONFIG, then the default; each applies to that one process. Setup passes the file on to the commands it registers in your agents, so they read the same login.

Supported agents

Setup only adds memry's own entries: everything else in those files is kept, and memry uninstall removes just what setup added. Your login token is never written to an agent's files.

How it works

• Login by email code. You sign in with a one-time code sent to your email; there is no password to remember.

• Agent integration. Setup registers a memry MCP server in each agent you select, which gives the agent the tools to save and search memories.

• Context at session start. Each new session loads the recent memories of the current project, so your agent picks up where you left off. Claude Code loads them with a hook (also on resumed and compacted sessions); the other agents are instructed to load them as their first step.

• Projects. By default, a repository's project is its directory name. To group several repositories into one project (for example, the backend and frontend of one product), add a .memry.json file at the root of each repository:

  {"project": "my-product"}

Usage metrics (opt-in)

memry sends no telemetry. If you want the memry dashboard to chart your Claude Code token usage and cost, turn it on explicitly:

memry usage enable    # off by default
memry usage status    # on or off as far as memry can see, and where to (never prints the token)
memry usage disable   # turn it off again

memry usage enable tells you what is sent, asks your memry server for a usage token (it can only send usage metrics) and sets Claude Code's OpenTelemetry metrics export in the env of ~/.claude/settings.json (or $CLAUDE_CONFIG_DIR/settings.json, a relative CLAUDE_CONFIG_DIR resolved against the current directory): CLAUDE_CODE_ENABLE_TELEMETRY, OTEL_METRICS_EXPORTER, OTEL_EXPORTER_OTLP_METRICS_PROTOCOL, OTEL_EXPORTER_OTLP_METRICS_ENDPOINT, OTEL_EXPORTER_OTLP_METRICS_HEADERS, and every OTEL_METRICS_INCLUDE_* gate set to false. Restart Claude Code to apply it. The settings file holds the usage token, so memry leaves it readable by your user only.

What is sent and what is kept

• Claude Code decides what it sends, as documented in its monitoring guide: metrics such as token usage, cost, sessions and lines of code, with identity attributes such as your email and organization id and metric-specific attributes. That list is not exhaustive and may grow with new Claude Code versions. memry turns off the optional attribute gates it can.

• The memry server is the guarantee. It keeps only the token and cost data points (the count or cost, the model, the token type, the query source and the time window), tied to your memry account, and discards everything else on receipt. It stores no raw attributes: only a keyed hash of them, to tell series apart.

• memry does not turn on Claude Code's logs or traces, so no prompts or code are sent through it.

What memry checks, and what it cannot see

Before asking for a token, enable checks the environment of the shell it runs in, your settings.json and Claude Code's file-based managed settings (managed-settings.json and managed-settings.d/*.json in /Library/Application Support/ClaudeCode on macOS, /etc/claude-code on Linux and WSL, C:\Program Files\ClaudeCode on Windows). It stops, whatever --force says, on any telemetry setting there that is not memry's own variable with memry's value: every OTEL_*, CLAUDE_CODE_ENABLE_TELEMETRY*, CLAUDE_CODE_ENHANCED_TELEMETRY*, ENABLE_ENHANCED_TELEMETRY*, CLAUDE_CODE_OTEL_* and *BETA_TRACING* variable and the otelHeadersHelper setting, except exporters set to none and variables you had already set to memry's value. A managed settings file it cannot read, or one that sets a policyHelper (memry never runs it), stops it too, as the effective policy cannot be determined. One of memry's own variables set in settings.json to another value stops it unless you pass --force. A plain http:// server that is not this machine needs --insecure, like memry setup.

memry cannot see per-session --settings files, what managed policy helpers compute, MDM profiles, the Windows registry, server-managed settings or the environment Claude Code is actually launched from. Any of them can change what Claude Code sends or where, so status reports what it finds "as far as memry can see". If you use mTLS, Claude Code may also present its client certificate (CLAUDE_CODE_CLIENT_CERT) to your memry server.

How memry keeps track

• memry records what it wrote under usage_metrics in the memry config (readable only by your user): the server, the settings file, each variable with its exact value, the usage token's id, and the variables you had already set to memry's value, which it never writes nor removes. A variable is memry's only while it has exactly the recorded value. disable, memry uninstall and memry delete-account remove just those, from the recorded file, drop an env left empty, and keep the record until that worked.

• memry revokes only the exact usage token it recorded, on the server it recorded, so your other machines' tokens keep working; revocation is best effort, and an unused token stops working after 30 days. enable for another server or CLAUDE_CONFIG_DIR stops until you disable the first one; memry setup to another server keeps the record.

• enable saves the new setup as pending before writing the settings, so an interrupted enable is settled by the next usage command. A failed enable changes nothing and the token in use keeps working.

• memry writes the settings only if the file is still, byte for byte, the one it checked; it compares again right before replacing it. A file system has no true compare-and-swap with other programs, so if another program writes settings.json at that same instant, the last writer wins. Usage commands, memry setup, memry uninstall and memry delete-account share a lock in memry/ under your user cache directory, which must be private to you, so they never overwrite each other's changes to the memry config. A lock is never taken over; after a crash, the error names the lock file to delete.

• If the memry config exists but cannot be read as a JSON object, the usage commands, memry uninstall and memry delete-account stop without changing anything.

Update

brew upgrade memry

On Windows, run winget upgrade Memry.Memry. Without a package manager, run the same go install again, or replace the binary with the one from a newer release. Your login and agent configuration keep working after an upgrade.

Uninstall

memry uninstall
brew uninstall memry

memry uninstall asks for confirmation, then removes memry from your agents and deletes your local login. Use memry uninstall --force to skip the confirmation. Without Homebrew, delete the memry binary instead of running brew uninstall. On Windows, run winget uninstall memry instead (memry uninstall prints the hint that fits your install).

Delete your account

memry delete-account

memry delete-account permanently deletes your memry account and all its memories on the server, then removes memry from your agents and deletes your local login, like memry uninstall. It cannot be undone, so it asks you to type your account email to confirm.

Your data

Your memories are stored on the memry server and tied to your account. memry uninstall revokes the token on this machine, so it can no longer access them. memry delete-account deletes them. Usage metrics are sent only if you turn them on (see Usage metrics). See the privacy policy (en español) for what memry stores and who handles it.

License

memry is released under the MIT license. See CHANGELOG.md for release notes and CONTRIBUTING.md for development and troubleshooting.