Agent Memory Hub

Provides AI agents with persistent, searchable long-term memory using a local JSON database and BM25 ranking for facts, preferences, and project details.

Category: Memory & Context
Version: 1.0.0
Type: MCP Server

Give any AI agent true long-term memory. Store facts, preferences, project details, and notes. Retrieve them with intelligent search. Never start from scratch again.

---

What It Does

agent-memory-hub is an MCP server that provides persistent, searchable memory for AI agents. It stores information between sessions using a local JSON database and retrieves it using BM25 full-text search with importance and recency weighting.

Think of it as a second brain for your agent — one that actually remembers who you are, what you're working on, and what you prefer.

---

When to Use

Situation	Action
Session start	Call get_relevant_context with the user's first message
User shares a preference	Call store_memory immediately
User references past work	Call search_memory before responding
User asks "do you remember…"	Call search_memory or get_relevant_context
End of important session	Call memory_summary to review what was captured
User corrects outdated info	Call update_memory

---

Tools Reference

store_memory(key, content, tags?, importance?)

Store any fact, preference, or note. Tags and importance are auto-detected.

Good keys: user_name, preferred_stack, project_deadline, api_key_note, client_requirement_1

store_memory("user_preferred_editor", "User uses VS Code with Vim keybindings")
→ Auto-tags: ["preference", "technical"]
→ Auto-importance: 6/10

---

search_memory(query, limit?, tags?)

BM25 ranked search across all memories. Weights importance and recency.

search_memory("code editor preferences")
→ Returns top matches with scores

---

get_relevant_context(user_query)

The power tool. Automatically surfaces the best memories for the current task. Call this at the start of every session.

get_relevant_context("Help me refactor the auth module")
→ TECHNICAL: preferred_stack, framework_version
→ PROJECT: auth_module_notes, project_deadline
→ PREFERENCE: user_coding_style

---

update_memory(key, new_content?, tags?, importance?)

Update any field of an existing memory without deleting it.

update_memory("project_deadline", "Deadline moved to June 20th")

---

list_memories(tags?, limit?, sort?)

Browse stored memories. Sort by recent, importance, or access count.

list_memories(tags=["project"], sort="importance", limit=10)

---

forget_memory(key)

Permanently remove a memory.

forget_memory("old_api_endpoint")

---

memory_summary()

High-level overview: count, top tags, most important, most accessed.

---

Example Agent Workflow

User: "Let's continue working on the SaaS dashboard"

Agent step 1: get_relevant_context("SaaS dashboard project")
→ Retrieves: project goals, tech stack, deadline, previous decisions

Agent step 2: [Responds with full context, no need to re-explain]

User: "By the way, we switched from Supabase to PlanetScale"

Agent step 3: store_memory("database_choice", "Using PlanetScale (switched from Supabase in May 2025)", importance=8)

[Next session — agent immediately knows this without being told again]

---

Installation

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "agent-memory-hub": {
      "command": "node",
      "args": ["/path/to/agent-memory-hub/build/index.js"]
    }
  }
}

Claude Code CLI

claude mcp add agent-memory-hub -- node /path/to/agent-memory-hub/build/index.js

Build from source

npm install && npm run build

---

Storage

• Default: ~/.agent-memory/memories.json
• Override: set AGENT_MEMORY_DIR environment variable
• Format: human-readable JSON, easy to backup

---

Auto-Tag Categories

preference · project · identity · technical · task · credential · note · person · config

---

Requirements

• Node.js 18+
• No API keys
• No external servers
• No Python or Docker