memento-mcp 0.3.0

The Autonomous Nervous System for AI Agents — local-first MCP middleware with temporal graph memory, active coercion, and proactive goal alignment.

Memento

The Autonomous Nervous System for AI Agents

---

Memento is a revolutionary, open-source middleware that acts as the "Autonomous Nervous System" for your AI agents (like Cursor, Claude Desktop, or Trae).

While most agentic memory systems rely on expensive, cloud-hosted graph databases, Memento empowers your local agents with a powerful, zero-cost, PageRank-optimized SQLite temporal graph and Reciprocal Rank Fusion (RRF) for perfect semantic retrieval.

But Memento goes Beyond Memory. It transforms your AI from a reactive assistant into a proactive, context-aware, and strictly aligned pair-programmer.

🌟 Enterprise-Grade Architecture

1. 🧠 Hybrid Search Engine (RRF)

A zero-cost, local-first temporal graph memory provider optimized for AI.

• Built on SQLite FTS5 (Full-Text Search) and Cosine Similarity (Vector Embeddings).
• Fuses exact keyword matches and semantic meaning using Reciprocal Rank Fusion (RRF).
• Write-Ahead Logging (WAL) enabled for extreme concurrency without database locking.
• Completely private and runs locally.

2. 🛡️ Active Coercion (The Code Immune System)

A deterministic, regex/AST-based engine that physically prevents the AI (or you) from introducing known anti-patterns.

• Pre-commit Hook Integration: Automatically blocks git commit if the code violates architectural rules.
• IDE Runtime Notifications: Sends push notifications directly to the AI if it generates bad code.
• 100% Deterministic: Zero LLM hallucinations during enforcement. Bypassable via // memento-override tokens.

3. 🎯 Tri-State Goal Enforcer

Keep your AI strictly aligned with your project's core objectives:

• Level 1 (Context Injection): Seamlessly injects active goals into the AI's context on every memory retrieval.
• Level 2 (Strict Mentor Checkpoint): Forces the AI to submit code or plans for a strict evaluation against the project's core goals.
• Level 3 (Proactive Autonomy): The AI is instructed via MCP to autonomously query Memento before writing any code.

Goal Enforcer MCP Tools:

Tool	Description
memento_set_goals	Set active goals (replace or append mode)
memento_list_goals	List goals with optional context and active-only filters
memento_check_goal_alignment	L2 gate — submit code/plans for strict goal evaluation
memento_configure_enforcement	Toggle L1/L2/L3 enforcement levels

4. 🕸️ Dynamic Workspace Router

Zero-config multi-tenant isolation. Memento automatically detects which project repository the AI is currently working on and routes the memory/database to the correct .memento/ folder. No more context bleeding between your Frontend and Backend projects.

---

🧬 The Seven Superpowers

Memento's cognitive layer goes beyond passive memory. Seven autonomous superpowers transform it into a self-improving, proactive system:

SP1: Auto-Consolidation

Automatically detects semantically similar memories and merges them into enriched, deduplicated entries. Uses cosine similarity clustering with sentence-level text fusion.

• memento_consolidate_memories — run a full consolidation cycle
• memento_toggle_consolidation_scheduler — start/stop background scheduler

SP2: KG Auto-Extraction

Automatically extracts entities and relationships from memories and populates a temporal knowledge graph using LLM analysis.

• memento_extract_kg — extract entities and triples from unprocessed memories
• memento_toggle_kg_extraction_scheduler — start/stop background scheduler

SP3: Relevance Tracking

Tracks memory access patterns with hit counting, temporal boosting, and exponential time decay. Frequently accessed recent memories rank higher.

• memento_get_relevance_stats — hot/cold distribution, hit counts, decay metrics
• memento_record_memory_hit — manually boost specific memories

SP4: Predictive Cache

Pre-warms an in-memory cache of related memories before starting work. Proactive context injection that anticipates what the AI will need.

• memento_warm_predictive_cache — warm cache with context text
• memento_get_predictive_cache_stats — hit rate, cache size, TTL info

SP5: Self-Evaluation Loop

Computes memory health scores (0-100) based on freshness, coverage, redundancy, and size. Identifies stale and orphan memories for cleanup.

• memento_get_quality_report — full quality report with health score
• memento_record_quality_evaluation — rate memory quality (0-1)
• memento_system_health — comprehensive system health dashboard
• memento_kg_health — knowledge graph entity/triple metrics

SP6: Cross-Workspace Sharing

Share memories between different Memento workspaces. Enables cross-project context flow with directional sync tracking.

• memento_share_memory_to_workspace — share a memory to another project
• memento_get_cross_workspace_stats — sync statistics

SP7: Real-Time Notifications

Proactive alerts about relevant context changes, memory events, and high-relevance discoveries. Configurable topics and confidence thresholds.

• memento_configure_notifications — enable/disable, set topics and confidence
• memento_get_pending_notifications — retrieve pending alerts
• memento_dismiss_notification — dismiss an alert

---

🚀 Quick Start

Choose your preferred installation method:

Option A: pip install (Recommended)

pip install memento-mcp

That's it. The memento-mcp and memento commands are now available globally.

Option B: uvx (Zero-install)

Run Memento instantly without installing anything permanently:

uvx memento-mcp

Option C: pip install from GitHub (Latest dev)

pip install git+https://github.com/JoyciAkira/memento.git

Option D: Clone for development

git clone https://github.com/JoyciAkira/memento.git
cd memento
uv sync

Verify any installation:

python -c "import memento; print(memento.__version__)"
memento-mcp --help
memento --help

📝 Running without OpenAI (offline / testing)

Set MEMENTO_EMBEDDING_BACKEND=none to disable embeddings entirely. Memento falls back to FTS5-only full-text search — no API key needed.

MEMENTO_EMBEDDING_BACKEND=none memento-mcp

---

🛠️ MCP Configuration (Cursor / Trae / Claude)

Add Memento to your mcp.json or IDE configuration. Thanks to the Dynamic Workspace Router, you only need to configure it once globally.

Config for pip / uvx install (Recommended)

No local clone needed. Just use the installed command directly:

{
  "mcpServers": {
    "memento": {
      "command": "memento-mcp",
      "env": {
        "OPENAI_API_KEY": "your-api-key-here",
        "OPENAI_BASE_URL": "https://api.openai.com/v1",
        "MEM0_MODEL": "openai/gpt-4o-mini",
        "MEM0_EMBEDDING_MODEL": "text-embedding-3-small"
      }
    }
  }
}

Config for local clone (development)

{
  "mcpServers": {
    "memento": {
      "command": "uv",
      "args": [
        "--directory",
        "/absolute/path/to/memento",
        "run",
        "memento-mcp"
      ],
      "env": {
        "OPENAI_API_KEY": "your-api-key-here",
        "OPENAI_BASE_URL": "https://api.openai.com/v1",
        "MEM0_MODEL": "openai/gpt-4o-mini",
        "MEM0_EMBEDDING_MODEL": "text-embedding-3-small"
      }
    }
  }
}

Environment variables

Variable	Description
OPENAI_API_KEY	Required for embeddings and goal checks
OPENAI_BASE_URL	Optional OpenAI-compatible endpoint
MEM0_MODEL	LLM used for cognitive features
MEM0_EMBEDDING_MODEL	Embeddings model used by the hybrid memory provider
MEMENTO_EMBEDDING_BACKEND	Set to none to disable embeddings (FTS5-only fallback)
MEMENTO_DIR	Workspace root used for routing .memento/ state
MEMENTO_UI	Enable local UI (1/true)
MEMENTO_UI_PORT	Local UI port (default 8089)

⌨️ CLI Usage

Memento also works directly from the terminal — no AI agent required.

# Auto-capture git context (branch, recent commits, diff stats) as a memory
memento capture --auto

# Save a free-form note
memento capture --text "Resolved the auth timeout by increasing JWT expiry to 1h"

# Combine auto context + custom note
memento capture --auto --text "Refactored the retry logic after the incident"

# Search your memories
memento search "how did I fix the promise bug"

# Show workspace status
memento status

The capture --auto command extracts the current git branch, last 5 commits, and staged/unstaged diff stats — saving a snapshot of your work context with zero friction. Useful in git hooks, CI scripts, or just as a quick terminal habit.

🧠 Using Memento (via MCP)

Memento exposes a suite of MCP tools, but the primary entrypoint is fully autonomous.

The Proactive Subconscious

The primary memento tool is configured with a Level 3 Proactive Autonomy prompt. You don't need to say "Memento, remember this". The AI is strictly instructed to query Memento before executing any task, formulating its own search queries to retrieve your architectural rules, past bugs, and context.

Managing Active Coercion

You can manage the Code Immune System directly via chat using the exposed MCP tools:

• memento_toggle_active_coercion
• memento_install_git_hooks
• memento_add_active_coercion_rule
• memento_list_active_coercion_rules

---

⚖️ License

Memento is released under the GNU Affero General Public License v3.0 (AGPL-3.0). This ensures that Memento remains free and open-source forever. If you modify Memento and offer it as a service over a network (e.g., as a Cloud SaaS), you must release your modified source code under the same AGPL-3.0 license.

See the LICENSE file for more details.