aimemo

English | 中文

Zero-dependency MCP memory server for AI agents — persistent, searchable, local-first, single binary.

$ claude "let's keep working on the payment service"

  ╭─ memory_context ──────────────────────────────────────────────────╮
  │ [project: payment-service]                                        │
  │                                                                   │
  │ Last session (3 days ago):                                        │
  │  • Stripe webhook signature verification — DONE                   │
  │  • Idempotency key refactor — IN PROGRESS                         │
  │  • Blocked: race condition in concurrent refund handler           │
  │                                                                   │
  │ Related: Redis connection pool, pkg/payments/refund.go            │
  ╰───────────────────────────────────────────────────────────────────╯

  Picking up where we left off. The race condition in the refund
  handler looks like a missing mutex around the in-flight map.
  Let me check pkg/payments/refund.go ...

  [... Claude works through the fix ...]

  ╭─ memory_store (journal) ──────────────────────────────────────────╮
  │ Resolved refund race — added sync.Mutex around inFlightRefunds.   │
  │ Tests passing. Next: load test with k6 at 500 rps.               │
  ╰───────────────────────────────────────────────────────────────────╯

🧠 Why aimemo

• No infra to babysit. Single Go binary. No Docker, no Node.js runtime, no cloud account, no API keys. brew install in 30 seconds.
• Memory stays with the project. Stored in .aimemo/ next to your code — commit it to git or add it to .gitignore. Switch branches; memory follows the directory.
• Claude picks up exactly where it left off. memory_context fires automatically on every session start. Claude sees what it was doing, what was blocked, what decisions were made. You stop repeating yourself.
• Full-text search that ranks correctly. FTS5 + BM25 scoring weighted by recency and access frequency. Relevant memories surface first; old noise fades naturally.
• Concurrent sessions, no corruption. SQLite WAL mode lets multiple Claude windows write simultaneously without locking each other out.
• You stay in control. Every tool Claude has, you have from the terminal. Inspect, edit, retract, export. Your memory is readable Markdown or JSON — never locked in a proprietary format.

⚡ Quick Start

# 1. Install
# Linux/macOS (one-line install):
curl -sSL https://raw.githubusercontent.com/MyAgentHubs/aimemo/main/install.sh | bash

# Or macOS via Homebrew:
brew install MyAgentHubs/tap/aimemo

# 2. Initialize memory for your project (run from project root)
aimemo init

# 3. Register with Claude Code
claude mcp add-json aimemo-memory '{"command":"aimemo","args":["serve"]}'

Restart Claude Code. On the next session, Claude will automatically load project context.

Quick Start for OpenClaw

If you're using OpenClaw skills, see the OpenClaw Integration section below for per-skill memory isolation.

🔧 How It Works

aimemo serve runs as a stdio MCP server; Claude Code manages the process lifecycle, so there is nothing to keep alive yourself. When Claude starts a session it calls memory_context to load relevant prior context; as it works it calls memory_store and memory_link to record decisions and relationships. You can call aimemo search, aimemo list, or aimemo get at any time to read or edit the same data from your terminal. Everything lives in a SQLite database inside .aimemo/, discovered by walking up from the current directory — the same way Git finds .git/.

🛠 MCP Tools

All tool schemas total under 2,000 tokens. Each call has a hard 5-second timeout — the server never stalls your session. Empty-state queries return in under 5 ms.

📋 CLI Reference

Setup

Memory

Journal

Inspect & Export

All commands accept --context <name> to target a named context (a separate .db file inside .aimemo/).

⚙️ Configuration

~/.aimemo/config.toml — global defaults, all optional:

[defaults]
context = "main"          # default context name
max_results = 20          # observations returned by memory_context

[scoring]
recency_weight = 0.7      # 0–1, weight of recency vs. access frequency

[server]
timeout_ms = 5000         # hard timeout on every MCP call
log_level = "warn"        # "debug" | "info" | "warn" | "error"

Per-project overrides live in .aimemo/config.toml in the project root — same keys, project values win over global values.

🤖 Claude Code Integration

Register the server once per machine:

claude mcp add-json aimemo-memory '{"command":"aimemo","args":["serve"]}'

Add the following to your project's CLAUDE.md so Claude knows memory is available and how to use it:

## Memory

This project uses aimemo for persistent memory across sessions.

- Call `memory_context` at the start of every session to load prior context.
- Call `memory_store` with `type: journal` before ending a session to record
  what was completed, what is still in progress, and any blockers.
- Use `memory_link` to connect related observations (e.g. a bug to its fix,
  a decision to its rationale).
- Do not store secrets, credentials, or PII.

🦞 OpenClaw Integration

aimemo solves OpenClaw's "remembers everything but understands none" problem with per-skill memory isolation and zero infrastructure.

Why aimemo for OpenClaw?

The Problem:

• OpenClaw's native Markdown memory gets worse the more you use it
• Skills share memory, causing cross-contamination
• Context compression loses important context

The Solution:

• ✅ Zero dependencies — Single Go binary, no Docker/Node.js/databases
• ✅ Per-skill isolation — Each skill gets its own memory database
• ✅ Actually works — BM25 search + importance scoring finds what matters
• ✅ Local-first — All data stays on your machine

vs Other Solutions:

5-Minute Setup

# 1. Install (Linux amd64/arm64)
curl -sSL https://raw.githubusercontent.com/MyAgentHubs/aimemo/main/install.sh | bash

# 2. Register MCP server with OpenClaw
claude mcp add-json aimemo-memory '{"command":"aimemo","args":["serve"]}'

# Or add to ~/.openclaw/openclaw.json:
# {
#   "mcpServers": {
#     "aimemo-memory": {
#       "command": "/usr/local/bin/aimemo",
#       "args": ["serve"]
#     }
#   }
# }

# 3. Initialize workspace memory
cd ~/.openclaw/workspace
aimemo init

# 4. Restart OpenClaw Gateway
# Linux: systemctl --user restart openclaw-gateway
# macOS: launchctl stop com.openclaw.gateway && launchctl start com.openclaw.gateway

Per-Skill Memory Isolation

Each skill gets its own isolated memory by using the context parameter: