MCP Memory Service — Use Cases, Install & Live Demo

Why use it

Key features

Vector + symbolic hybrid recall — fewer false positives than pure embedding stores
Auto-built knowledge graph linking entities, decisions, and projects
Background consolidation: dedupes, summarizes, and surfaces drift
REST API alongside MCP — same store, two clients
SQLite-first (zero infra); pluggable to Postgres / pgvector for scale

Live Demo

What it looks like in practice

mcp-memory-service.replay ▶ ready

0/0

Install

Pick your client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "mcp-memory-service": {
      "command": "uvx",
      "args": [
        "mcp-memory-service"
      ]
    }
  }
}

Open Claude Desktop → Settings → Developer → Edit Config. Restart after saving.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "mcp-memory-service": {
      "command": "uvx",
      "args": [
        "mcp-memory-service"
      ]
    }
  }
}

Cursor uses the same mcpServers schema as Claude Desktop. Project config wins over global.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "mcp-memory-service": {
      "command": "uvx",
      "args": [
        "mcp-memory-service"
      ]
    }
  }
}

Click the MCP Servers icon in the Cline sidebar, then "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "mcp-memory-service": {
      "command": "uvx",
      "args": [
        "mcp-memory-service"
      ]
    }
  }
}

Same shape as Claude Desktop. Restart Windsurf to pick up changes.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "mcp-memory-service",
      "command": "uvx",
      "args": [
        "mcp-memory-service"
      ]
    }
  ]
}

Continue uses an array of server objects rather than a map.

~/.config/zed/settings.json

{
  "context_servers": {
    "mcp-memory-service": {
      "command": {
        "path": "uvx",
        "args": [
          "mcp-memory-service"
        ]
      }
    }
  }
}

Add to context_servers. Zed hot-reloads on save.

claude mcp add mcp-memory-service -- uvx mcp-memory-service

One-liner. Verify with claude mcp list. Remove with claude mcp remove.

Use Cases

Real-world ways to use MCP Memory Service

Keep Claude in sync across a multi-week coding project

👤 Developers on long-running projects ⏱ ~15 min beginner

When to use: Every session you spend 10 minutes reminding Claude of project conventions, decisions, and TODOs.

Prerequisites

Memory service running — uvx mcp-memory-service (defaults to ~/.mcp-memory.db)

Flow

Seed core context

Use mcp-memory-service. Store these as project-level memory: 'use bun not npm', 'all DB writes go through src/db/repo.ts', 'no class components in src/ui'.✓ Copied

→ 3 memories stored with project tag
Use in next session

Recall any project-level rules before suggesting changes to src/ui/Form.tsx.✓ Copied

→ Relevant rules surfaced; suggestion respects them
Capture new decisions

Save: 'switched from React Query to SWR on 2026-04-22 because of Suspense compat'.✓ Copied

→ Memory stored with date; will surface next time topic comes up

Outcome: Claude treats your project like a continuous engagement, not 100 cold starts.

Pitfalls

Memory grows unbounded; recall starts pulling old/irrelevant stuff — Run consolidate weekly; tag memories so you can scope recall to recent + this-project

Combine with: filesystem

Give a personal assistant agent a persistent profile of you

👤 People building daily-driver Claude agents ⏱ ~10 min beginner

When to use: Your daily agent should know your preferences, recurring meetings, and people without re-asking.

Flow

Onboard the profile

Use mcp-memory-service. Store: my timezone (Asia/Tokyo), my partner's name, my work hours (10–7), recurring 1:1s.✓ Copied

→ Profile stored with type=profile
Recall on the fly

Schedule 'lunch w/ Alice' for next Tuesday — pick a time that respects my work hours.✓ Copied

→ Suggestion uses your hours from memory; no re-asking

Outcome: Less repeated context-setting per task.

Pitfalls

Sensitive details (medical, financial) stored unencrypted — Use a separate DB path (encrypted FS or SQLCipher build); or simply don't store medical/financial in this layer

Combine with: google-workspace-mcp

Build a research knowledge graph as you read papers

👤 Researchers, analysts, deep-dive readers ⏱ ~60 min intermediate

When to use: You're reading 30 papers and want to track entities, claims, and contradictions.

Flow

Ingest as you go

For each paper I link, extract: authors, key claims, methods, datasets — store as linked nodes.✓ Copied

→ Node count growing; links visible
Find contradictions

Across stored papers, find claims that contradict. List with sources.✓ Copied

→ Pairs of conflicting claims with citations
Generate a survey

Draft a 2-page survey with the dominant claim threads + contradictions noted.✓ Copied

→ Markdown survey with sourced claims

Outcome: A real synthesis, not just notes.

Pitfalls

Extraction quality varies — some claims are noise — Periodically prune low-confidence nodes; the consolidation job helps but isn't magic

Combine with: filesystem

Share project decision log across teammates' agents

👤 Small teams using AI dev agents ⏱ ~30 min advanced

When to use: Three people on the team use Claude Code; nobody knows what decisions the others' agents have already made.

Prerequisites

Hosted instance reachable by all — Run with --host 0.0.0.0 on a shared dev box; or deploy Docker image

Flow

Point team configs at shared service

Update each teammate's MCP config to use the shared URL with separate auth tokens.✓ Copied

→ All agents read/write the same store
Convention: tag decisions

Convention: anything stored as type=decision is team-visible; type=personal is namespaced.✓ Copied

→ Tags applied; recall scoped accordingly

Outcome: Team-level memory without a wiki.

Pitfalls

Personal notes accidentally tagged as team-visible — Default to type=personal; require explicit type=decision opt-in for team notes

Combine with: github

Combinations

Pair with other MCPs for X10 leverage

mcp-memory-service + filesystem

Anchor memory in repo files — link memory_id ↔ file path

When I edit src/db/repo.ts, recall any decisions referencing it before suggesting changes.✓ Copied

mcp-memory-service + github

Auto-store PR review summaries as decision memory

After every merged PR, store the design rationale + reviewer concerns as type=decision tagged with the repo.✓ Copied

Tools

What this MCP exposes

Tool	Inputs	When to call	Cost
store_memory	content, tags?, type?, metadata?	On any new fact / decision worth keeping	0
recall	query, top_k?, tags?, since?	First step before answering — don't reinvent context	0
graph_neighbors	node_id, depth?	Explore related entities/decisions	0
consolidate	scope?	Weekly housekeeping; merges duplicates	LLM tokens for summarization
delete_memory	memory_id	Right-to-be-forgotten or wrong-data correction	0

Cost & Limits

What this costs to run

API quota: None — all local
Tokens per call: Recall returns 200–2000 tokens depending on top_k
Monetary: Free; consolidation uses local or BYO LLM
Tip: Cap top_k=5 for most recall calls; let consolidation handle dedupe instead of overstuffing prompts

Security

Permissions, secrets, blast radius

Minimum scopes: filesystem write to db path

Credential storage: DB at ~/.mcp-memory.db; encrypt the FS or use SQLCipher build for sensitive data

Data egress: None by default; consolidation can call out if you wire it to an LLM

Never grant: World-readable DB on shared hosts

PII flows in via memories — apply data-retention policy at the application layer

Troubleshooting

Common errors and fixes

Recall returns nothing

Check tags filter; embeddings index may need rebuild after schema changes

Verify: Look at tool result count and the tags you pass

DB locked errors under concurrent agents

Switch to Postgres backend or enable SQLite WAL mode (default in recent versions)

Verify: PRAGMA journal_mode;

Consolidation runs forever

Scope it: consolidate(scope='last_week') instead of full DB

Memories full of off-topic chat

Be stricter at store time — only call store_memory on facts/decisions, not casual chat

Verify: Audit recent stores; delete cruft

Alternatives

MCP Memory Service vs others

Alternative	When to use it instead	Tradeoff
claude-mem-skill	You want session-bridging compression, not a long-term knowledge store	claude-mem is per-session continuity; this is true persistent memory + graph
Mem0	You want a hosted memory layer with SDKs	Costs money at scale; data leaves your machine
Letta (formerly MemGPT)	You want full agent runtime with built-in memory, not just a memory layer	Heavier; opinionated about agent architecture

More

Resources

📖 Read the official README on GitHub

🐙 Browse open issues

🔍 Browse all 400+ MCP servers and Skills