Graphmem
Shared, local memory for coding agents. Keep project decisions across sessions and tools with SQLite storage, local semantic search, and linked entities. Claude Code, Codex, OpenCode, pi, and other MCP clients can share the same store.
Optional code tools find definitions, outline files, list imports, and compare symbols between Git revisions.
Install
npm install --global @sonic182/graphmem
gmem-install
gmem version
The npm package contains launchers. gmem-install explicitly downloads and
verifies the matching CPU binary, or reuses a working native gmem on your
PATH. Package installation itself downloads no binary. After upgrading the
package, run gmem-install again; separately installed binaries are not replaced.
MCP clients that launch servers with npx can run npx -y @sonic182/graphmem mcp.
The graphmem command downloads and verifies the binary on its first run; gmem
never downloads anything on its own.
Alternatively, download your platform's archive from
GitHub Releases, verify
it against SHA256SUMS, and put the extracted executable on your PATH.
Linux releases require glibc 2.39 and OpenSSL 3. For other systems or CUDA,
build from source.
Use it with your coding agent
Plugins include MCP registration, skills, and guidance to recall context before work and save durable decisions afterward.
Claude Code and Codex need gmem and Node.js on PATH. In Codex, trust the
Graphmem hooks in /hooks and start a new thread. Pi offers binary setup on first
startup; OpenCode prints an installer command if needed.
See plugin setup and troubleshooting.
For another MCP client, configure it to launch your installed binary:
{
"mcpServers": {
"graphmem": {
"command": "/absolute/path/to/gmem",
"args": ["mcp"]
}
}
}
CLI
gmem remember "Use nextest for integration tests" \
--type convention --scope repo:/absolute/path/to/project
gmem search "integration tests"
gmem tokens --over-limit # memories the embedding model truncates
gmem tui
gmem mcp
Memories default to the current Git repository; git must be on PATH.
Outside a repository, the default scope is global. Reads always include global
memories; explicitly select other projects with --scope (CLI) or scopes (MCP).
Writes are limited to the current repository and global, including updates and
deletions. The MCP list_scopes tool discovers scopes and their write permissions;
list pages through the memories of a selection.
Data lives in ~/.graphmem; set GRAPHMEM_HOME to use another directory.
CLI reference · MCP tool reference
Code navigation (optional)
Release binaries include these tools; source builds need --features code.
gmem code find CodeService
gmem code outline src/application/code.rs
gmem code imports src/application/code.rs
gmem code diff origin/master
These tools work on Git checkouts without an embedding model. They index
definitions, not call sites or references; use ast-grep for those.
Set GRAPHMEM_CODE=off to disable code tools.
Options and supported languages · Benchmark results and limitations
Configuration
No config file is required. To customize embeddings, retrieval, or code indexing,
create ~/.graphmem/config.toml. See the configuration reference.
Build
See source builds and CUDA setup.
Development
just verify
Runs formatting checks, compilation, Clippy, and tests. Design · Schema evolution
To preview the documentation site, run hugo server --source site.
See site setup.
License
MIT.
