By anortham
Cross-client MCP memory with checkpoints, recall, briefs, and standups for AI-assisted development
Use when the user asks for progress against current project direction, wants a status check on the active brief, or needs evidence-backed assessment across briefs, checkpoints, and docs/plans
Use when starting multi-session work, capturing approved project direction, updating goals or constraints, or when the user asks for durable strategic context that should persist across sessions
Save developer context to Goldfish memory — checkpoint at meaningful milestones, not after every action
Use when returning to a project after time away, switching harnesses, or handing work off to another agent, to produce a structured session-resumption summary from the active brief, recent checkpoints, and git delta.
Use when starting a new session, after context loss, searching for prior work, or when the user asks what happened previously and needs Goldfish memory restored
Admin access level
Server config contains admin-level keywords
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
An evidence ledger for AI coding sessions. Checkpoints capture what changed and why; briefs hold durable strategic direction; recall pulls both back when the next session needs context. Everything lives as markdown in your repo, so it travels with the code, diffs in PRs, and outlasts any single harness.
Goldfish is a cross-client MCP memory system. Claude Code gets the fullest adapter today, with plugin installation and slash-command skills. Codex Desktop and OpenCode can discover repo-local Goldfish skills from .agents/skills, and VS Code with GitHub Copilot can use the MCP server plus repo instructions.
Version 7.2.0 -- Recall gains type and tags filters, search now indexes checkpoint type, and corrupt registry / brief / lock files are handled defensively instead of wiping data or stalling. See CHANGELOG.md for details.
Coding harnesses already plan, summarize, and recover from compaction. What they don't do is keep a durable record of why a project moved the way it did, in a place the next session (or the next harness) can read.
Goldfish is git for intent: a source-controlled, harness-agnostic ledger of decisions, milestones, and direction. Three MCP tools (checkpoint, recall, brief) and six skills, with markdown as the source of truth.
Prerequisites: Bun runtime (v1.0+)
Start by cloning the repository and installing dependencies:
git clone https://github.com/anortham/goldfish.git
cd goldfish
bun install
Claude Code is the fullest adapter today. You get MCP tools and slash-command skills (/checkpoint, /recall, /brief, /brief-status, /handoff, /standup).
Install from the marketplace:
# Add the Goldfish repository as a plugin marketplace
/plugin marketplace add anortham/goldfish
# Install the plugin for your user
/plugin install goldfish@goldfish
# Or scope it to the current project
/plugin install goldfish@goldfish --scope project
Install from a local clone:
claude plugin install /path/to/goldfish
For development, load the plugin from the local directory each time:
claude --plugin-dir /path/to/goldfish
Once the plugin is loaded, Goldfish works through manual invocation and agent-driven calls:
/recall (or let the agent call recall()) to restore recent checkpoints and the active brief/checkpoint at meaningful milestones/brief when goals, constraints, or success criteria should survive the sessionCodex shares MCP configuration between the CLI and the IDE extension through ~/.codex/config.toml, and it also discovers repo-local skills from .agents/skills.
Codex Desktop does not send MCP roots. If you want Goldfish bound to the current repo, the reliable setup is a project-local .codex/config.toml in that repo so you can pass GOLDFISH_WORKSPACE for that project.
Add Goldfish to a trusted project-local .codex/config.toml:
[mcp_servers.goldfish]
command = "bun"
args = ["run", "/absolute/path/to/goldfish/src/server.ts"]
cwd = "/absolute/path/to/your/project"
env = { GOLDFISH_WORKSPACE = "/absolute/path/to/your/project" }
You can put Goldfish in ~/.codex/config.toml too, but that pins GOLDFISH_WORKSPACE to one repo. For Codex Desktop across multiple repos, keep the server entry in each project's .codex/config.toml.
Goldfish skills in .agents/skills are discovered automatically when you launch Codex inside the repository.
OpenCode loads local MCP servers from opencode.json and can also discover repo-local skills from .agents/skills.
Add Goldfish to your opencode.json:
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"goldfish": {
"type": "local",
"command": ["bun", "run", "/absolute/path/to/goldfish/src/server.ts"],
"enabled": true
}
}
}
OpenCode walks up the repository and loads matching .agents/skills/*/SKILL.md, so the checked-in Goldfish skills are available without extra copying.
VS Code supports project-level MCP config in .vscode/mcp.json, supports the full MCP feature set, and can pair Goldfish with repo instructions for better memory habits.
mkdir -p .vscode
Create .vscode/mcp.json:
{
"servers": {
"Goldfish": {
"type": "stdio",
"command": "bun",
"args": ["run", "/absolute/path/to/goldfish/src/server.ts"]
}
}
}
GOLDFISH_WORKSPACE is optional in VS Code now that Goldfish can resolve the active workspace from MCP roots. Keep it as an override if you want to pin Goldfish to a different root or you run in a client that does not provide roots:
Code intelligence server: search, navigation, and refactoring across 34 languages
Miller-powered development workflow with subagent-driven execution, inline review, and Miller-first codebase orientation.
Agent memory persistence - remember and recall across sessions
Composite workflows that orchestrate Julie (code intelligence), Sherpa (workflow guidance), and Goldfish (persistent memory) for powerful development patterns. Includes TDD Powerhouse, Bug Detective, Smart Session Start, and more.
Local SQLite-backed code intelligence, content search, and workspace freshness for coding agents.
npx claudepluginhub anortham/goldfish --plugin goldfishAuto-capture high-signal coding context into memctl memory
Persistent memory for AI coding agents. Survives across sessions and compactions.
Persistent memory system for AI coding sessions — cross-tool memory sharing with 6-dimensional hybrid search
Persistent memory for Claude Code. Capture work across sessions and recall relevant context.
OpenLTM — Long-Term Memory for AI coding agents: semantic search, context injection, session learning
Curated persistent memory for Claude Code. Write gate prevents bloat — only behavior-changing facts get saved. Tiered architecture: daily logs, structured registers, and auto-loaded working memory.