Agents Remember MCP

📖 Current docs: https://foxfire1st.github.io/agents-remember/
🤖 Machine-readable summary: https://foxfire1st.github.io/agents-remember/llms.txt
📦 Repository: https://github.com/Foxfire1st/agents-remember
Note: caches and search snippets may serve an outdated copy of this README — the docs site above is canonical and always current.

agents-remember-mcp is the installable Model Context Protocol server for Agents Remember. It lets an MCP-capable coding harness call Agents Remember operations from the host instead of asking the model to edit or execute coordinator scripts directly.

Quickstart

Setup is agent-driven. Ask your agent to:

1. Copy the harness package — Pick your harness guide under docs/install, copy that harness's native starter files from the source repo into the workspace, then render the copied package. The render-starter script is a convenience: it infers the workspace root from the copied harness folder and fills the copied package's path, repository, and hook-command placeholders from a single --repo list such as --repo my-app shared-lib. You can also do those replacements by hand. These packages include the harness-visible skills, hooks/rules/instructions, MCP settings templates, and local renderers.

2. Wire the MCP server — Register Agents Remember MCP with uvx and the copied settings file:

   uvx agents-remember-mcp@latest --config /absolute/path/to/agents-remember-settings.json
   

   Use the agents-remember-settings.json path from the copied harness package. Then restart the harness once so it loads the MCP server, native skills, and package hooks/rules/instructions.

3. Onboard your project — Invoke the copied skill c-13-install-and-onboard. It runs or verifies runtime_install(), asks whether to scaffold a new memory repo or use an existing one, bootstraps onboarding when needed, and starts provider indexing when providers are enabled.

That is the normal first-run path. skills_install() remains available as a maintenance/manual MCP tool, but the starter packages already provide the initial skills and harness files. Your only required first-run restart is after copying the harness package and wiring the MCP server.

Requirements

• Python 3.11 or newer (the package declares requires-python >=3.11; on a multi-version host, uvx selects a compatible interpreter automatically).
• an MCP-capable coding harness
• uv (for uvx) or pip
• Git for repository and memory ledger operations (configure user.name / user.email; without them, memory/worktree commits fall back to a placeholder identity so work can proceed).
• Docker (running) when provider tools are enabled. The grepai embedder runs Ollama as a Docker container and pulls an embedding model (nomic-embed-text) on first setup — no host Ollama install is required.

Claude Code hooks do not require jq. Older starter packages used a jq one-liner to encode hook output; the current starter packages use Python renderers and Python hook scripts.

Install And Run

The simplest path is uvx, which fetches and runs the server on demand — no manual virtualenv or PATH setup:

uvx agents-remember-mcp --config /absolute/path/to/agents-remember-settings.json

Or install with pip and use the console command:

python -m pip install agents-remember-mcp
agents-remember-mcp --config /absolute/path/to/agents-remember-settings.json

The config path must be absolute, the settings file must live outside the ar-coordination/ runtime folder, and it should live under your harness's registration folder in an mcp/ subdirectory (see Settings file location). The server reads this file only at startup — if you edit it later (enable providers, add repos, change timeoutCaps), restart the harness for the change to take effect; run server_info() to confirm what the running server actually loaded.

Settings

A minimal starter agents-remember-settings.json (your agent can fill this in):

{
  "version": 1,
  "coordinationRoot": "/absolute/path/to/ar-coordination",
  "workspaceRoot": "/absolute/path/to/workspace",
  "repositories": {
    "<your-repo-name>": {}
  },
  "providers": {
    "codegraphcontext-code": {},
    "grepai-memory": {}
  }
}

coordinationRoot is where the runtime and memory repos live (populated by the runtime_install MCP tool); default it to <workspace>/ar-coordination/ — inside the workspace, never the user's home directory. workspaceRoot is the workspace itself (it holds your code repos). The c-13-install-and-onboard skill treats the workspace as the first assumption for every install location and shows each resolved default for you to accept or override, so placement is never silent or guessed. List each repo you want Agents Remember to manage under repositories. Omit or empty the providers block if you do not want the Docker-backed providers. Full field reference: settings-json.md.

Upgrading? timeoutCaps.providerSeconds was renamed to providerSetupSeconds. The old key is rejected with a fail-loud ConfigError at startup, so rename it in any existing settings file. providerSetupSeconds caps only provider image build / dependency install; indexing and database seed/clone are never time-capped. A cap value of 0 means unlimited.

Settings file location

Place the settings file where the copied starter package expects it. Keep it under the harness registration folder, not loose in the workspace root and not inside ar-coordination/.

See your harness page under docs/install/ for the exact registration folder.

Harness Setup

Register the MCP server with your harness by pointing it at uvx (or the installed console command) and the absolute settings path:

{
  "command": "uvx",
  "args": [
    "agents-remember-mcp",
    "--config",
    "/absolute/path/to/agents-remember-settings.json"
  ]
}

After installing or changing the MCP server registration, restart the harness so it reloads the server and discovers the tool list. Use the starter package for your harness whenever possible; it already carries the matching skills, hooks/rules/instructions, and settings template.

Per-harness setup pages

Harnesses differ in where settings and skills go — so follow your harness's page, don't guess:

One flat folder per skill. The copied starter package already includes the skills in the harness-native skill root. skills_install() remains available for manual maintenance and non-package setups; it copies packaged skills into a skill root as <skill-root>/<name>/ (matching the skill's lowercase frontmatter name).

Install Order And First Operations

With starter packages, the strict first-run order is package + MCP wiring → one harness restart → runtime and onboarding.

server_info()                      # confirm resolved roots / allowed providers
runtime_install(dry_run=true)      # preview, then apply:
runtime_install(dry_run=false)     # scaffold coordinator; build provider images if enabled
context_packet(repo_id="<repo-id>", include_providers=true)

The copied c-13-install-and-onboard skill owns this post-restart phase. It runs or verifies runtime_install(), sets up the memory repo, bootstraps onboarding, and starts provider indexing (provider_watchers(action="start")) when providers are enabled.

Why this order:

1. Harness-native files first. Skills, hooks/rules/instructions, and MCP settings are loaded by the harness, so the copied starter package must be in place before restart.
2. Runtime scaffolding after restart. The MCP server must be loaded before the agent can run runtime_install(). The runtime tool creates the coordinator directory and records the provider-runner integrity manifest.
3. Providers last. They are heavy (Docker, plus Ollama for grepai), per-repo, and optional. Note the split: runtime_install() builds provider images when install_provider_deps=true, but indexing only starts later via c-13-install-and-onboard. Pass install_provider_deps=false to refresh scaffold/docs without rebuilding images or disturbing running watchers; pass no_cache=true to force a from-scratch image rebuild (it otherwise skips images whose tag already exists). If providers report degraded, check that Docker is running and (for grepai) that the Ollama model pulled, then provider_watchers(action="refresh"); provider_diagnostics() shows the gap.

Troubleshooting

uvx can't find a just-published version. PyPI's simple index (what uvx resolves against) lags a few minutes behind a release, so uvx agents-remember-mcp==X.Y.Z may briefly fail with "no version found" right after that version is published. Wait 2–5 minutes and retry, run uvx --refresh … to bypass uv's cache, or drop the ==X.Y.Z pin to take the latest the index currently serves.

Providers report degraded and indexing/search returns nothing. Both providers need Docker running; grepai additionally needs its Ollama container and embedding model. Check and recover:

• Docker: docker ps — if the daemon is down, start it (sudo systemctl start docker on Linux, or Docker Desktop), then provider_watchers(action="refresh").
• grepai/Ollama: docker logs ar-grepai-ollama and docker exec ar-grepai-ollama ollama list to confirm the model is present. The model (nomic-embed-text) is pulled automatically on first setup; if that timed out, re-run runtime_install() or pull it manually with docker exec ar-grepai-ollama ollama pull nomic-embed-text.
• provider_diagnostics() shows the precise failing resource (backend, embedder, watcher) for either provider.

Providers are optional — core by-path memory and onboarding work without them, so you can defer this and the rest of setup continues.

Memory/worktree commits and git identity. Closeout and carryover operations commit to the memory repo and ledger, so git needs an author identity. Configure your own with git config --global user.name "…" and git config --global user.email "…". If none is set, Agents Remember writes a repo-local placeholder (Agents Remember <agents-remember@example.invalid>) so work can still proceed — commits just won't carry your identity until you set it.

Tool Surface

The server exposes tools for:

• startup context and drift checks
• runtime and skill installation
• memory initialization, memory quality checks, and route index refresh
• provider status, watcher control, GrepAI search, and CodeGraphContext queries
• worktree-backed closeout and task workflows
• benchmark preparation and execution (opt-in; see the note below)

Provider tools only work when the MCP settings enable the provider and the required Docker services are available. Full tool list: MCP Tool Reference.

Benchmark execution is opt-in and runs untrusted code. The codex_benchmark_prepare and codex_benchmark_run MCP tools are refused unless the MCP settings set "benchmarksEnabled": true. A real run (dry_run=false) clones third-party repositories and executes the Codex CLI against them. codex_sandbox defaults to Codex's own default sandbox; pass "danger-full-access" only for trusted local runs — it grants the benchmark agent full host access.

More

• Project README
• Getting Started
• Settings Reference