Production-Hardened MCP Memory Server — Hybrid Search + Resilience for AI Agents

The only MCP memory server with circuit breaker, SLO tracking, and BM25+FAISS hybrid search. AI agents forget everything between sessions. MindCore Memory gives them persistent, searchable, production-grade memory — with 118/118 tests passing and full CI/CD.

⭐ If this project helps your AI remember, a star means the world to us.

Quick Start

# 1. Install
pip install mindcore-memory

# 2. Launch (stdio mode — works with any MCP client)
mindcore-memory

# 3. Your AI agent remembers across sessions

{
  "mcpServers": {
    "mindcore-memory": {
      "command": "python",
      "args": ["-m", "mindcore_memory.server"],
      "env": { "MINDCORE_MEMORY_PATH": "~/.mindcore/memory" }
    }
  }
}

pip install mindcore-memory[semantic]
# Enables FAISS embeddings for hybrid BM25+semantic search

Why MindCore — vs the Competition

MindCore is the only MCP memory server designed for production workloads from day one. Circuit breaker protects against embedding service failures. Retry with exponential backoff handles transient errors. SLO tracking alerts you before users notice. Metrics export for your monitoring stack. Every other server assumes nothing fails — MindCore doesn't.

Unique: 3D Boundary Balance Algorithm

MindCore is not just a memory store — it's a cognitive boundary engine. Every stored memory is automatically evaluated through a 4-dimensional scoring system based on the 正反公式 (Forward/Reverse Formula):

BND_score = 0.28·TRJ(Trajectory) + 0.28·EVO(Evolution) + 0.28·COG(Cognition) + 0.16·BALANCE

• Forward cycle: TRJ → BND → EVO → COG → BND (each step draws a boundary, each boundary is growth)
• Reverse chain: Chaos → Unknown → Risk → Harm → Death (2+ linked triggers → auto 50% score penalty)
• 3D balance: Variance across TRJ/EVO/COG penalizes lopsided memories (pure data dumps without insight)
• No LLM calls: Pure algorithmic evaluation using keyword patterns, regex, and statistical variance

from mindcore_memory import BNDManager
bnd = BNDManager()
result = bnd.evaluate("基于之前修复, 理解到根因, 改进后提升30%", importance=4)
# → TRJ:0.63  EVO:0.54  COG:0.61  BALANCE:0.98  BND:0.75  ACCEPTED

📖 Full algorithm documentation

No other MCP memory server does this. BND transforms memory storage from a passive data dump into an active cognitive filter — rejecting noise, flagging risk chains, and ensuring only structured, growth-oriented knowledge enters the version chain.

Production Features

Resilience Layer

• Circuit Breaker: CLOSED → OPEN → HALF_OPEN state machine. Protects FAISS/embedding operations from cascading failure.
• Retry: Exponential backoff with jitter. Transient errors retry automatically, permanent errors fail fast.
• Input Validation: Server-level sanitization against injection attacks.

Observability Layer

• SLO Tracking: P95/P99 latency targets for all 6 operations. Violations logged and exported.
• Prometheus /metrics: Zero-dependency Prometheus-compatible collector. Drop-in for any monitoring stack.

Data Layer

• Encryption: Optional Fernet encryption at rest (mindcore-memory[encrypt]).
• Deduplication: Exact-match merge — identical memory updates importance/confidence instead of storing duplicates.
• Smart Eviction: Low-importance memory pruning with atomic disk sync. No zombie memories.

Core Tools

Memory (6 tools)

Boundary & Deduction (3 tools) 🆕

Search formula: score = BM25(40%) + FAISS(50%) + importance(5%) + recency(5%)

When FAISS embeddings are unavailable, automatically falls back to BM25-only keyword search.

Architecture

┌───────────────────┐     MCP JSON-RPC      ┌────────────────────────────┐
│  AI Client         │ ◄──────────────────► │  MindCore Memory           │
│  (Claude/Cursor)   │     stdio / HTTP     │  MCP Server                │
└───────────────────┘                       └──────────┬─────────────────┘
                                                       │
                                            ┌──────────▼─────────────────┐
                                            │  Memory Engine             │
                                            │  ┌──────────────────────┐  │
                                            │  │ Hybrid Search        │  │
                                            │  │  BM25 (keyword) 40%  │  │
                                            │  │  FAISS (semantic)50%│  │
                                            │  │  importance        5%│  │
                                            │  │  recency           5%│  │
                                            │  └──────────────────────┘  │
                                            │  ┌──────────────────────┐  │
                                            │  │ Resilience           │  │
                                            │  │  Circuit Breaker     │  │
                                            │  │  Retry + Backoff     │  │
                                            │  │  SLO Tracking        │  │
                                            │  └──────────────────────┘  │
                                            └──────────┬─────────────────┘
                                                       │
                                            ┌──────────▼─────────────────┐
                                            │  Storage                   │
                                            │  JSONL (append)            │
                                            │  + FAISS index (IVF > 500) │
                                            │  + Fernet encrypt (opt)    │
                                            └────────────────────────────┘

• Embedded: No PostgreSQL, Redis, or external services needed. One binary, local JSONL + FAISS.
• IVF Index: FAISS inverted file index activates at 500+ memories for O(√N) search.
• MCP Native: Full MCP protocol over stdio and HTTP transports.

Available On

Full Comparison