mem-agent-mcp

mem-agent-mcp is a Python MCP (Model Context Protocol) server that connects an “obsidian-like” local Markdown memory store (user.md + entities/*.md) to LLM clients (e.g., Claude Desktop, Lm Studio, Claude Code) for memory retrieval and memory-related tool operations. It also documents memory import/connector workflows (chatgpt, notion, nuclino, github, google-docs) and ways to run the model backend locally (MLX on macOS or vLLM/OpenAI-compatible via LiteLLM proxy).

Evaluated Mar 30, 2026 (22d ago)

Repo ↗ Ai Ml mcp memory obsidian-style rag local-first python tool-integration file-system connectors vllm mlx llm-clients

⚙ Agent Friendliness

/ 100

Can an agent use this?

🔒 Security

/ 100

Is it safe for agents?

⚡ Reliability

/ 100

Does it work consistently?

Score Breakdown

⚙ Agent Friendliness

MCP Quality

Documentation

Error Messages

Auth Simplicity

Rate Limits

🔒 Security

TLS Enforcement

Auth Strength

Scope Granularity

Dep. Hygiene

Secret Handling

Strengths: local-first design with user-managed tokens; Google Drive scope guidance is explicit (drive.readonly), and GitHub token scopes are described (public_repo vs repo). Risks/unknowns: MCP server authentication/authorization is not documented; MCP over HTTP could be exposed without access control. Token handling and logging practices are not described; ensure tokens are not logged and that file permissions for the memory directory are restricted. TLS is only mentioned implicitly (HTTP endpoint example); no explicit 'HTTPS required' guidance is provided.

⚡ Reliability

Uptime/SLA

Version Stability

Breaking Changes

Error Recovery

Best When

You want local-first private memory with an MCP client, and you can run/point to a local model server (MLX/vLLM or an OpenAI-compatible proxy).

Avoid When

You need strict security boundaries for remote access or you cannot control tokens, local storage permissions, and logs; also avoid if you need standardized HTTP APIs beyond MCP.

Use Cases

• Connect an Obsidian-style local Markdown memory directory to an MCP-capable client
• RAG-like memory Q&A over imported conversations and entities
• Maintaining/updating personal or team memory from exports (ChatGPT/Notion/Nuclino)
• Live memory augmentation from GitHub repositories and Google Docs via Drive APIs
• Applying retrieval-time filters provided through the prompt (<filter> tags)
• Using MCP over stdio or HTTP for different client environments

Not For

• Public multi-tenant production deployments without proper network isolation/authn/authz
• Handling highly sensitive data without a threat model for local file access and logs
• Environments that cannot run a local model backend or an OpenAI-compatible proxy
• Use cases requiring a formal, documented REST/OpenAPI developer platform (this is primarily MCP/file-based)

Interface

REST API

GraphQL

gRPC

MCP Server

Yes ↗

SDK

Webhooks

Authentication

Methods: Environment-variable configuration for model proxy/backend connectivity (VLLM_HOST/VLLM_PORT) Connector tokens for GitHub (TOKEN) Google Drive OAuth access tokens (ACCESS TOKEN / access token guidance) Google Docs connector uses Drive API access token Interactive authentication setup via memory wizard (tokens/scopes described)

OAuth: No Scopes: Yes

No explicit MCP-auth mechanism is described in the README. Connector access relies on user-provided tokens (GitHub classic token; Google Drive OAuth scopes like drive.readonly). If MCP-over-HTTP is exposed (e.g., via ngrok), MCP itself appears to be unauthenticated unless the deployment adds external auth.

Pricing

Free tier: No

Requires CC: No

Self-hosted/local model usage; costs depend on model/backend hardware or any OpenAI-compatible proxy you run (e.g., LiteLLM/OpenRouter). No hosted pricing described.

Agent Metadata

Pagination

none

Idempotent

False

Retry Guidance

Not documented

Known Gotchas

⚠ MCP server functionality appears tied to local filesystem layout (memory/user.md and memory/entities/*.md); incorrect directory structure may break retrieval/tool behavior.
⚠ On ARM64 Linux, vLLM may not be installed by default; recommended path is OpenAI-compatible via LiteLLM proxy, which can surprise agents expecting a local vLLM binary.
⚠ Some workflows (filters, connectors) modify local .filters and memory content; lack of described idempotency means repeated runs could duplicate/overwrite depending on connector implementation.
⚠ If MCP-over-HTTP is publicly exposed (e.g., via ngrok), the README does not describe authentication/authorization safeguards for MCP endpoints.

Alternatives

Other MCP memory/tool servers (generic file+search MCP implementations) Building a custom MCP server over your own vector store + document store Using a managed memory/RAG service with explicit enterprise security controls Standalone retrieval services (e.g., local search + reranking) without MCP

Full Evaluation Report

Comprehensive deep-dive: security analysis, reliability audit, agent experience review, cost modeling, competitive positioning, and improvement roadmap for mem-agent-mcp.

AI-powered analysis · PDF + markdown · Delivered within 30 minutes

$99

Package Brief

Quick verdict, integration guide, cost projections, gotchas with workarounds, and alternatives comparison.

Delivered within 10 minutes

Score Monitoring

Get alerted when this package's AF, security, or reliability scores change significantly. Stay ahead of regressions.

Continuous monitoring

$3/mo

API endpoint ↗ Agent guide ↗ Report inaccuracy

Scores are editorial opinions as of 2026-03-30.