mcp-memory-service

Persistent Shared Memory for AI Agent Pipelines

Open-source memory backend for multi-agent systems. Agents store decisions, share causal knowledge graphs, and retrieve context in 5ms — without cloud lock-in or API costs.

Works with LangGraph · CrewAI · AutoGen · any HTTP client · Claude Desktop

🎬 See It in Action

Watch the Web Dashboard Walkthrough on YouTube — Semantic search, tag browser, document ingestion, analytics, quality scoring, and API docs in under 2 minutes.

🌐 Works with claude.ai (Browser)

Unlike desktop-only MCP servers, mcp-memory-service supports Remote MCP for native claude.ai integration.

What this means:

• ✅ Use persistent memory directly in your browser (no Claude Desktop required)
• ✅ Works on any device (laptop, tablet, phone)
• ✅ Enterprise-ready (OAuth 2.0 + HTTPS + CORS)
• ✅ Self-hosted OR cloud-hosted (your choice)

5-Minute Setup:

# 1. Start server with Remote MCP enabled
MCP_STREAMABLE_HTTP_MODE=1 \
MCP_SSE_HOST=0.0.0.0 \
MCP_SSE_PORT=8765 \
MCP_OAUTH_ENABLED=true \
python -m mcp_memory_service.server
 
# 2. Expose via Cloudflare Tunnel (or your own HTTPS setup)
cloudflared tunnel --url http://localhost:8765
# → Outputs: https://random-name.trycloudflare.com
 
# 3. In claude.ai: Settings → Connectors → Add Connector
# Paste the URL: https://random-name.trycloudflare.com/mcp
# OAuth flow will handle authentication automatically

Production Setup: See Remote MCP Setup Guide for Let's Encrypt, nginx, and firewall configuration. Step-by-Step Tutorial: Blog: 5-Minute claude.ai Setup | Wiki Guide

Why Agents Need This

Key capabilities for agent pipelines:

• Framework-agnostic REST API — 15 endpoints, no MCP client library needed
• Knowledge graph — agents share causal chains, not just facts
• X-Agent-ID header — auto-tag memories by agent identity for scoped retrieval
• conversation_id — bypass deduplication for incremental conversation storage
• SSE events — real-time notifications when any agent stores or deletes a memory
• Embeddings run locally via ONNX — memory never leaves your infrastructure

Agent Quick Start

pip install mcp-memory-service
MCP_ALLOW_ANONYMOUS_ACCESS=true memory server --http
# REST API running at http://localhost:8000

import httpx
 
BASE_URL = "http://localhost:8000"
 
# Store — auto-tag with X-Agent-ID header
async with httpx.AsyncClient() as client:
    await client.post(f"{BASE_URL}/api/memories", json={
        "content": "API rate limit is 100 req/min",
        "tags": ["api", "limits"],
    }, headers={"X-Agent-ID": "researcher"})
    # Stored with tags: ["api", "limits", "agent:researcher"]
 
# Search — scope to a specific agent
    results = await client.post(f"{BASE_URL}/api/memories/search", json={
        "query": "API rate limits",
        "tags": ["agent:researcher"],
    })
    print(results.json()["memories"])

Framework-specific guides: docs/agents/

Real-World: Multi-Agent Cluster with Shared Memory

"After I work with one of the cluster agents on something I want my local agent to know about, the cluster agent adds a special tag to the memory entry that my local agent recognizes as a message from a cluster agent. So they end up using it as a comms bridge — and it's pretty delightful." — @jeremykoerber, issue #591

A 5-agent openclaw cluster uses mcp-memory-service as shared state and as an inter-agent messaging bus — without any custom protocol. Cluster agents tag memories with a sentinel like msg:cluster, and the local agent filters on that tag to receive cross-cluster signals. The memory service becomes the coordination layer with zero additional infrastructure.

# Cluster agent stores a learning and flags it for the local agent
await client.post(f"{BASE_URL}/api/memories", json={
    "content": "Rate limit on provider X is 50 RPM — switch to provider Y after 40",
    "tags": ["api", "limits", "msg:cluster"],       # sentinel tag
}, headers={"X-Agent-ID": "cluster-agent-3"})
 
# Local agent polls for cluster messages
results = await client.post(f"{BASE_URL}/api/memories/search", json={
    "query": "messages from cluster",
    "tags": ["msg:cluster"],
})

This pattern — tags as inter-agent signals — emerges naturally from the tagging system and requires no additional infrastructure.

Real-World: Self-Hosted Docker Stack with Cloudflare Tunnel

"The quality of life that session-independent memory adds to AI workflows is immense. File-based memory demands constant discipline. Semantic recall from a live database doesn't. Storing data on my own hardware while making it remotely accessible across platforms turned out to be a feature I didn't know I needed." — @PL-Peter, discussion #602

A production-tested self-hosted deployment using Docker containers behind a Cloudflare tunnel, with AuthMCP Gateway handling authentication:

Security best practices for this setup:

• Use Cloudflare ZeroTrust with subnet-based access control (e.g., allow Anthropic subnets + your own IPs)
• Add Client IP Address Filtering to all Cloudflare API tokens (Dashboard → My Profile → API Tokens → Edit → Client IP Address Filtering) to limit abuse if a token leaks
• If using IPv6, include your IPv6 /64 network in the allowlist (Python prefers IPv6 by default)
• Set MCP_OAUTH_ACCESS_TOKEN_EXPIRE_MINUTES=1440 to extend OAuth tokens to 24 hours (refresh tokens not yet supported)
• Consider an auth proxy like AuthMCP or mcp-auth-proxy for robust session management

Comparison with Alternatives

vs. Commercial Memory APIs

vs. MCP-Native Alternatives

MemPalace (~20k ⭐) is a strong MCP-native alternative worth knowing about.

Why the benchmark gap? MemPalace stores each conversation as a single unit (session-level). LongMemEval asks "which session contains the answer?" — a question that session-level storage answers structurally. mcp-memory-service defaults to turn-level storage (one entry per message), which enables fine-grained retrieval ("what exactly did the user say about X?") but spreads a session's signal across many entries. Using memory_store_session (session-level ingestion, added in v10.35.0) brings our score to 86.0% R@5 — closing the gap significantly. The remaining difference is primarily due to MemPalace's larger embedding model.

¹ 100% result uses optional LLM reranking (~500 API calls) and includes a partially tuned test set. Clean held-out score: 98.4% R@5.

Stop Re-Explaining Your Project to AI Every Session

Your AI assistant forgets everything when you start a new chat. After 50 tool uses, context explodes to 500k+ tokens—Claude slows down, you restart, and now it remembers nothing. You spend 10 minutes re-explaining your architecture. Again.

MCP Memory Service solves this.

It automatically captures your project context, architecture decisions, and code patterns. When you start fresh sessions, your AI already knows everything—no re-explaining, no context loss, no wasted time.

🎥 2-Minute Video Demo

⚡ Works With Your Favorite AI Tools

🤖 Agent Frameworks (REST API)

LangGraph · CrewAI · AutoGen · Any HTTP Client · OpenClaw/Nanobot · Custom Pipelines

🖥️ CLI & Terminal AI (MCP)

Claude Code · Gemini CLI · Gemini Code Assist · OpenCode · Codex CLI · Goose · Aider · GitHub Copilot CLI · Amp · Continue · Zed · Cody

🎨 Desktop & IDE (MCP)

Claude Desktop · VS Code · Cursor · Windsurf · Kilo Code · Raycast · JetBrains · Replit · Sourcegraph · Qodo

💬 Chat Interfaces (MCP)

ChatGPT (Developer Mode) · claude.ai (Remote MCP via HTTPS)

Works seamlessly with any MCP-compatible client or HTTP client - whether you're building agent pipelines, coding in the terminal, IDE, or browser.

💡 NEW: ChatGPT now supports MCP! Enable Developer Mode to connect your memory service directly. See setup guide →

🚀 Get Started in 60 Seconds

Not sure which setup fits your needs? See the Setup Guide — a decision tree walks you to the right path in under a minute.

1. Install:

pip install mcp-memory-service

2. Configure your AI client:

Add to your config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "memory": {
      "command": "memory",
      "args": ["server"]
    }
  }
}

Restart Claude Desktop. Your AI now remembers everything across sessions.

claude mcp add memory -- memory server

Restart Claude Code. Memory tools will appear automatically.

No local installation required on the client — works directly in your browser:

# 1. Start server with Remote MCP
MCP_STREAMABLE_HTTP_MODE=1 python -m mcp_memory_service.server
 
# 2. Expose publicly (Cloudflare Tunnel)
cloudflared tunnel --url http://localhost:8765
 
# 3. Add connector in claude.ai Settings → Connectors with the tunnel URL

See Remote MCP Setup Guide for production deployment with Let's Encrypt, nginx, and Docker.

For production deployments, team collaboration, or cloud sync:

git clone https://github.com/doobidoo/mcp-memory-service.git
cd mcp-memory-service
python scripts/installation/install.py

Choose from:

• SQLite (local, fast, single-user)
• Cloudflare (cloud, multi-device sync)
• Hybrid (best of both: 5ms local + background cloud sync)

💡 Why You Need This

The Problem

The Solution

Result: Zero re-explaining. Zero context loss. Just continuous, intelligent collaboration.

🌐 SHODH Ecosystem Compatibility

MCP Memory Service is fully compatible with the SHODH Unified Memory API Specification v1.0.0, enabling seamless interoperability across the SHODH ecosystem.

Compatible Implementations

Unified Schema Support

All SHODH implementations share the same memory schema:

• ✅ Emotional Metadata: emotion, emotional_valence, emotional_arousal
• ✅ Episodic Memory: episode_id, sequence_number, preceding_memory_id
• ✅ Source Tracking: source_type, credibility
• ✅ Quality Scoring: quality_score, access_count, last_accessed_at

Interoperability Example: Export memories from mcp-memory-service → Import to shodh-cloudflare → Sync across devices → Full fidelity preservation of emotional_valence, episode_id, and all spec fields.

✨ Quick Start Features

🧠 Persistent Memory – Context survives across sessions with semantic search 🔍 Smart Retrieval – Finds relevant context automatically using AI embeddings ⚡ 5ms Speed – Instant context injection, no latency 🔄 Multi-Client – Works across 20+ AI applications ☁️ Cloud Sync – Optional Cloudflare backend for team collaboration 🔒 Privacy-First – Local-first, you control your data 📊 Web Dashboard – Visualize and manage memories at http://localhost:8000 🧬 Knowledge Graph – Interactive D3.js visualization of memory relationships 🆕

🖥️ Dashboard Preview (v9.3.0)

8 Dashboard Tabs: Dashboard • Search • Browse • Documents • Manage • Analytics • Quality (NEW) • API Docs

📖 See Web Dashboard Guide for complete documentation.

Latest Release: v10.35.0 (April 8, 2026)

feat: session-level memory ingestion — LongMemEval R@5 86.0% (+5.6% vs turn-level)

What's New:

• memory_store_session MCP tool: Stores a full conversation as a single memory unit — all turns concatenated as [role] content, stored with memory_type=session and auto-tagged session:<id>.
• POST /api/sessions HTTP endpoint: REST endpoint for session-level ingestion mirroring the MCP tool.
• LongMemEval session-mode results: R@5 86.0% (+5.6% vs turn-level), with biggest gains in multi-session (+15.2%) and temporal-reasoning (+10.6%) categories.
• --ingestion-mode session|turn|both flag for LongMemEval benchmark for direct strategy comparison.
• session and conversation_turn memory types added to the ontology.
• 1,537 tests passing (+17 new: 10 handler + 7 HTTP endpoint tests).

Previous Releases:

• v10.34.0 - feat: LongMemEval benchmark — R@5 80.4%, R@10 90.4%, NDCG@10 82.2%, MRR 89.1% (PR #665, 1,520 tests)
• v10.33.0 - refactor: eliminate event-loop blocking + fix silent conflict data loss in SQLite storage (PR #663, 1,520 tests)
• v10.32.0 - feat: transport health endpoint + configurable timeouts + optional DCR registration key protection (community PRs #656, #657, 1,520 tests)
• v10.31.2 - fix: storage consistency, error handling, and upload progress — _safe_json_loads consistency, non-JSON error handling, upload progress tracking (community PRs #648, #649, #650, 1,503 tests)
• v10.31.1 - fix: tombstone blocks re-insertion after delete of same content (#644) — _purge_tombstone() before INSERT (1,521 tests)
• v10.31.0 - feat: Harvest Evolution (P4) + Sync-in-Async Refactoring — harvest dedup via update_memory_versioned(), asyncio.to_thread() in _execute_with_retry (1,520 tests)
• v10.30.0 - feat: Memory Evolution (P1+P2+P3) — non-destructive versioned updates, staleness scoring, conflict detection + resolution (1,514 tests)
• v10.29.1 - fix: clean up orphaned graph edges on memory deletion — cascade edge removal in delete/delete_by_tag/delete_by_tags + periodic orphan pruning in consolidation
• v10.29.0 - feat(harvest): LLM-based classification via Groq (Phase 2, #628) — memory_harvest supports use_llm=true for higher-precision category labels via _GroqClassifierBridge
• v10.28.5 - Bug fix: MCP_ALLOW_ANONYMOUS_ACCESS=true now respected in the dashboard (anonymous users granted read+write scope)
• v10.28.4 - Security patch: cryptography>=46.0.6 (CVE-2026-34073), serialize-javascript>=7.0.5 (CVE-2026-34043), CodeQL cleanup
• v10.28.3 - HTTP MCP endpoint fix: accept 'content' as alias for 'query' so Claude Code HTTP transport returns results
• v10.28.2 - Relationship inference tuning: 93.5% typed labels vs 0.5% before + German language support
• v10.28.1 - Harvest false-positive fix: skip system prompts, skill outputs, and long injected content (3 new tests)
• v10.28.0 - Session harvest tool (memory_harvest): extract learnings from Claude Code transcripts + security dependency updates (#614-#616)
• v10.27.0 - External embedding compatibility fix (missing index field, community PR #612) + Docker/Cloudflare deployment docs
• v10.26.8 - 6 bug fixes in consolidation, embeddings, and memory types (#603-#608)
• v10.26.7 - Cloudflare D1 fresh-database schema initialization fix (issue #600), community contribution by @Lyt060814
• v10.26.6 - Security patch: authlib>=1.6.9, PyJWT>=2.12.0, pypdf>=6.9.1 (5 Dependabot alerts: 1 critical, 3 high, 1 medium)
• v10.26.5 - Security patch: black dev dependency bumped to >=26.3.1 (GHSA-3936-cmfr-pm3m, CVE-2026-32274, path traversal)
• v10.26.4 - FTS5 hybrid search fix on upgrade + dashboard auth lifecycle fixes (9 bugs)
• v10.26.3 - Dashboard metadata display fixes + quality scorer resilience (Groq 429 fallback chain, empty-query absolute prompt)
• v10.26.2 - OAuth public PKCE client fix (token exchange 500 error, issue #576) + automated CHANGELOG housekeeping
• v10.26.1 - Hybrid backend correctly reported in MCP health checks (HealthCheckFactory structural detection fix for wrapped/delegated backends, issue #570)
• v10.26.0 - Credentials tab + Settings restructure + Sync Owner selector in dashboard; MCP_HYBRID_SYNC_OWNER=http recommended for hybrid mode
• v10.25.3 - Patch release: stdio handshake timeout cap, syntax fixes, hybrid sync fix, dashboard version badge fix
• v10.25.2 - Patch fix: update_and_restart.sh health check reads status field instead of removed version field
• v10.25.1 - Security: CORS wildcard default changed to localhost-only, soft-delete leak in search_by_tag_chronological() fixed (GHSA-g9rg-8vq5-mpwm)
• v10.25.0 - Embedding migration script, 5 soft-delete leak fixes, cosine distance formula fix, substring tag matching fix, O(n²) association sampling fix — 23 new tests, 1,420 total

Full version history: CHANGELOG.md | Older versions (v10.22.0 and earlier) | All Releases

Migration to v9.0.0

⚡ TL;DR: No manual migration needed - upgrades happen automatically!

Breaking Changes:

• Memory Type Ontology: Legacy types auto-migrate to new taxonomy (task→observation, note→observation)
• Asymmetric Relationships: Directed edges only (no longer bidirectional)

Migration Process:

1. Stop your MCP server
2. Update to latest version (git pull or pip install --upgrade mcp-memory-service)
3. Restart server - automatic migrations run on startup:
   • Database schema migrations (009, 010)
   • Memory type soft-validation (legacy types → observation)
   • No tag migration needed (backward compatible)

Safety: Migrations are idempotent and safe to re-run

Breaking Changes

1. Memory Type Ontology

What Changed:

• Legacy memory types (task, note, standard) are deprecated
• New formal taxonomy: 5 base types (observation, decision, learning, error, pattern) with 21 subtypes
• Type validation now defaults to 'observation' for invalid types (soft validation)

Migration Process: ✅ Automatic - No manual action required!

When you restart the server with v9.0.0:

• Invalid memory types are automatically soft-validated to 'observation'
• Database schema updates run automatically
• Existing memories continue to work without modification

New Memory Types:

• observation: General observations, facts, and discoveries
• decision: Decisions and planning
• learning: Learnings and insights
• error: Errors and failures
• pattern: Patterns and trends

Backward Compatibility:

• Existing memories will be auto-migrated (task→observation, note→observation, standard→observation)
• Invalid types default to 'observation' (no errors thrown)

2. Asymmetric Relationships

What Changed:

• Asymmetric relationships (causes, fixes, supports, follows) now store only directed edges
• Symmetric relationships (related, contradicts) continue storing bidirectional edges
• Database migration (010) removes incorrect reverse edges

Migration Required: No action needed - database migration runs automatically on startup.

Code Changes Required: If your code expects bidirectional storage for asymmetric relationships:

# OLD (will no longer work):
# Asymmetric relationships were stored bidirectionally
result = storage.find_connected(memory_id, relationship_type="causes")
 
# NEW (correct approach):
# Use direction parameter for asymmetric relationships
result = storage.find_connected(
    memory_id,
    relationship_type="causes",
    direction="both"  # Explicit direction required for asymmetric types
)

Relationship Types:

• Asymmetric: causes, fixes, supports, follows (A→B ≠ B→A)
• Symmetric: related, contradicts (A↔B)

Retrieval Benchmarks

Three benchmarks measure retrieval quality (all-MiniLM-L6-v2, 384d embeddings, zero LLM API calls):

LongMemEval (500 questions, ~45–62 distractor sessions per question):

DevBench (practical developer workflow queries):

LoCoMo (ACL 2024 long-term conversational memory):

Run benchmarks: python scripts/benchmarks/benchmark_longmemeval.py, python scripts/benchmarks/benchmark_devbench.py, python scripts/benchmarks/benchmark_locomo.py

Performance Improvements

• ontology validation: 97.5x faster (module-level caching)
• Type lookups: 35.9x faster (cached reverse maps)
• Tag validation: 47.3% faster (eliminated double parsing)

Testing

• 829/914 tests passing (90.7%)
• 80 new ontology tests with 100% backward compatibility
• All API/HTTP integration tests passing

Support

If you encounter issues during migration:

• Check Troubleshooting Guide
• Review CHANGELOG.md for detailed changes
• Open an issue: https://github.com/doobidoo/mcp-memory-service/issues

📚 Documentation & Resources

• Agent Integration Guides 🆕 – LangGraph, CrewAI, AutoGen, HTTP generic
• Remote MCP Setup (claude.ai) 🆕 – Browser integration via HTTPS + OAuth
• Setup Guide – Decision tree + step-by-step paths for all use cases
• Configuration Guide – Backend options and customization
• Architecture Overview – How it works under the hood
• Team Setup Guide – OAuth and cloud collaboration
• Knowledge Graph Dashboard 🆕 – Interactive graph visualization guide
• Troubleshooting – Common issues and solutions
• API Reference – Programmatic usage
• Wiki – Complete documentation
• – AI-powered documentation assistant
• MCP Starter Kit – Build your own MCP server using the patterns from this project

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Quick Development Setup:

git clone https://github.com/doobidoo/mcp-memory-service.git
cd mcp-memory-service
pip install -e .  # Editable install
pytest tests/      # Run test suite