agent-memory

Deprecated: use jcyamacho/memo for new durable memory setups. memo uses the same store format, so existing agent-memory users only need to point memo at the same store with MEMO_CONFIG_DIR. For the default agent-memory location, run MEMO_CONFIG_DIR=~/.config/agent-memory memo .... If this repo already uses AGENT_MEMORY_STORE_PATH, use that value as MEMO_CONFIG_DIR.

Persistent memory for MCP-powered coding agents.

agent-memory is a stdio MCP server that gives your LLM durable memory backed by Markdown files on disk. It helps your agent remember preferences, project context, and prior decisions across sessions.

It exposes four tools:

• remember: save facts, decisions, preferences, and project context
• review: load workspace and global memories sorted by most recently updated
• revise: update an existing memory when its content changes or when it should become global
• forget: delete up to 50 memories that are no longer relevant

Quick Start

Claude CLI:

claude mcp add --scope user agent-memory -- npx -y @jcyamacho/agent-memory

Codex CLI:

codex mcp add agent-memory -- npx -y @jcyamacho/agent-memory

OpenCode:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "agent-memory": {
      "type": "local",
      "command": ["npx", "-y", "@jcyamacho/agent-memory"]
    }
  }
}

Install the Agent Skill

Install the optional Agent Skill to teach supported coding agents how to create small, atomic memories and maintain them conservatively:

npx skills add jcyamacho/agent-memory --skill agent-memory -g

The skill complements the MCP server instructions with a detailed durability gate, scoping rules, and guidance for choosing between remember, revise, forget, and no operation. It does not change the MCP tools or storage format.

Load Memories at Session Start (Hooks)

Some clients defer MCP tool loading until the model decides it needs a tool. That can prevent proactive review calls at session start. A session-start hook guarantees memories are loaded before the first prompt.

The CLI prints the same <memories> XML as the MCP review tool:

npx -y @jcyamacho/agent-memory review

Use --workspace <path> to override the default workspace (process.cwd()). Hook commands run with the project directory as cwd, so the default is usually enough.

Claude Code (~/.claude/settings.json or .claude/settings.json):

{
  "hooks": {
    "SessionStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "npx -y @jcyamacho/agent-memory review"
          }
        ]
      }
    ]
  }
}

Codex CLI (~/.codex/hooks.json):

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "startup|resume|clear|compact",
        "hooks": [
          {
            "type": "command",
            "command": "npx -y @jcyamacho/agent-memory review"
          }
        ]
      }
    ]
  }
}

Notes:

• Claude Code caps hook stdout at 10,000 characters. Larger output is written to a file the model can read.
• npx cold-cache latency can delay session start. If the package is installed locally, call the agent-memory bin directly.
• Claude Code users can alternatively set "alwaysLoad": true on the agent-memory server in .mcp.json (v2.1.121+) to prioritize this MCP by making its tools visible from the first turn. This leaves the decision to call review under model control.

Optional LLM Instructions

Use these fallback instructions when the Agent Skill is not installed. The server instructions and tool descriptions already cover most behavior. If a session-start hook already injected memories, the prompt avoids loading them a second time:

## Agent Memory

- Use the `agent-memory` MCP as the only durable memory system when its tools are
  available.
- If workspace memories are not already present in context, call `review` from
  the `agent-memory` MCP at session start. Continue through every page while
  `has_more` is true.
- During the session, use `remember`, `revise`, and `forget` from the
  `agent-memory` MCP to keep memories accurate.
- Pass `workspace` for project-scoped memory. Omit it only for facts that truly
  apply across projects. Promote project-scoped memory to global only when that
  is actually true.
- Remember durable preferences, confirmed approaches, and decisions with
  reasoning that would be hard to recover later. Revise memories when facts
  change and forget them when they are no longer relevant.
- Do not store secrets, temporary task state, or facts obvious from current
  code, files, or git history.

Mutating Tool Output

remember and revise return the full affected memory as XML with updated_at and scope information so clients that hide tool-call arguments can still see what changed.

forget accepts an ids array containing 1 to 50 memory IDs. It trims IDs, ignores duplicates, and deletes up to five memories concurrently. Deletion is best-effort: one failure does not stop the remaining memories. The returned <forget_results> preserves request order and contains each deleted memory with deleted="true", failed IDs and statuses, and summary counts.

How Review Works

review requires a workspace and returns memories saved in that workspace plus global memories (saved without a workspace), sorted by most recently updated. Results are paginated. Pass page to load older memories.

When you save a memory from a git worktree, agent-memory stores the main repo root as the workspace. review applies the same normalization to incoming workspace queries so linked worktrees still match repo-scoped memories exactly.

Configuration

Store Location

By default, memories are stored under:

~/.config/agent-memory/

Override it with:

AGENT_MEMORY_STORE_PATH=/absolute/path/to/agent-memory

The store layout is:

<store>/
  globals/<memory-id>.md
  workspaces/<encoded-workspace>/<memory-id>.md

Set AGENT_MEMORY_STORE_PATH when you want to:

• keep memory in a project-specific location
• share a memory store across multiple clients
• keep the Markdown files somewhere easier to back up or inspect

License

MIT