MCP 服务器

qualixar/superlocalmemory

Persistent AI memory MCP server with 4-channel retrieval (semantic, BM25, entity graph, temporal). 74.8% on LoCoMo benchmark with zero cloud dependency. Works with Claude, Cursor, VS Code Copilot, and 17+ AI tools. EU AI Act compliant. MIT license.

README

<h1 align="center">SuperLocalMemory V3</h1> The first local-only AI memory to break 74% retrieval on LoCoMo. No cloud. No APIs. No data leaves your machine.

<code>+16pp vs Mem0 (zero cloud)</code> · <code>85% Open-Domain (best of any system)</code> · <code>EU AI Act Ready</code>

Why SuperLocalMemory?

Every major AI memory system — Mem0, Zep, Letta, EverMemOS — sends your data to cloud LLMs for core operations. That means latency on every query, cost on every interaction, and after August 2, 2026, a compliance problem under the EU AI Act.

SuperLocalMemory V3 takes a different approach: mathematics instead of cloud compute. Three techniques from differential geometry, algebraic topology, and stochastic analysis replace the work that other systems need LLMs to do — similarity scoring, contradiction detection, and lifecycle management. The result is an agent memory that runs entirely on your machine, on CPU, with no API keys, and still outperforms funded alternatives.

The numbers (evaluated on LoCoMo, the standard long-conversation memory benchmark):

System	Score	Cloud Required	Open Source	Funding
EverMemOS	92.3%	Yes	No	—
Hindsight	89.6%	Yes	No	—
SLM V3 Mode C	87.7%	Optional	Yes (MIT)	$0
Zep v3	85.2%	Yes	Deprecated	$35M
SLM V3 Mode A	74.8%	No	Yes (MIT)	$0
Mem0	64.2%	Yes	Partial	$24M

Mode A scores 74.8% with zero cloud dependency — outperforming Mem0 by 16 percentage points without a single API call. On open-domain questions, Mode A scores 85.0% — the highest of any system in the evaluation, including cloud-powered ones. Mode C reaches 87.7%, matching enterprise cloud systems.

Mathematical layers contribute +12.7 percentage points on average across 6 conversations (n=832 questions), with up to +19.9pp on the most challenging dialogues. This isn't more compute — it's better math.

Upgrading from V2 (2.8.6)? V3 is a complete architectural reinvention — new mathematical engine, new retrieval pipeline, new storage schema. Your existing data is preserved but requires migration. After installing V3, run slm migrate to upgrade your data. Read the Migration Guide before upgrading. Backup is created automatically.

Quick Start

Install via npm (recommended)

npm install -g superlocalmemory
slm setup     # Choose mode (A/B/C)
slm warmup    # Pre-download embedding model (~500MB, optional)

Install via pip

pip install superlocalmemory

First Use

slm remember "Alice works at Google as a Staff Engineer"
slm recall "What does Alice do?"
slm status

MCP Integration (Claude, Cursor, Windsurf, VS Code, etc.)

{
  "mcpServers": {
    "superlocalmemory": {
      "command": "slm",
      "args": ["mcp"]
    }
  }
}

24 MCP tools available. Works with Claude Code, Cursor, Windsurf, VS Code Copilot, Continue, Cody, ChatGPT Desktop, Gemini CLI, JetBrains, Zed, and 17+ AI tools.

Three Operating Modes

Mode	What	Cloud?	EU AI Act	Best For
A	Local Guardian	None	Compliant	Privacy-first, air-gapped, enterprise
B	Smart Local	Local only (Ollama)	Compliant	Better answers, data stays local
C	Full Power	Cloud LLM	Partial	Maximum accuracy, research

slm mode a   # Zero-cloud (default)
slm mode b   # Local Ollama
slm mode c   # Cloud LLM

Mode A is the only agent memory that operates with zero cloud dependency while achieving competitive retrieval accuracy on a standard benchmark. All data stays on your device. No API keys. No GPU. Runs on 2 vCPUs + 4GB RAM.

Architecture

Query  ──►  Strategy Classifier  ──►  4 Parallel Channels:
                                       ├── Semantic (Fisher-Rao geodesic distance)
                                       ├── BM25 (keyword matching)
                                       ├── Entity Graph (spreading activation, 3 hops)
                                       └── Temporal (date-aware retrieval)
                                                    │
                                       RRF Fusion (k=60)
                                                    │
                                       Scene Expansion + Bridge Discovery
                                                    │
                                       Cross-Encoder Reranking
                                                    │
                                       ◄── Top-K Results with channel scores

Mathematical Foundations

Three novel contributions replace cloud LLM dependency with mathematical guarantees:

Fisher-Rao Retrieval Metric — Similarity scoring derived from the Fisher information structure of diagonal Gaussian families. Graduated ramp from cosine to geodesic distance over the first 10 accesses. The first application of information geometry to agent memory retrieval.
Sheaf Cohomology for Consistency — Algebraic topology detects contradictions by computing coboundary norms on the knowledge graph. The first algebraic guarantee for contradiction detection in agent memory.
Riemannian Langevin Lifecycle — Memory positions evolve on the Poincare ball via discretized Langevin SDE. Frequently accessed memories stay active; neglected memories self-archive. No hardcoded thresholds.

These three layers collectively yield +12.7pp average improvement over the engineering-only baseline, with the Fisher metric alone contributing +10.8pp on the hardest conversations.

Benchmarks

Evaluated on LoCoMo — 10 multi-session conversations, 1,986 total questions, 4 scored categories.

Mode A (Zero-Cloud, 10 Conversations, 1,276 Questions)

Category	Score	vs. Mem0 (64.2%)
Single-Hop	72.0%	+3.0pp
Multi-Hop	70.3%	+8.6pp
Temporal	80.0%	+21.7pp
Open-Domain	85.0%	+35.0pp
Aggregate	74.8%	+10.6pp

Mode A achieves 85.0% on open-domain questions — the highest of any system in the evaluation, including cloud-powered ones.

Math Layer Impact (6 Conversations, n=832)

Conversation	With Math	Without	Delta
Easiest	78.5%	71.2%	+7.3pp
Hardest	64.2%	44.3%	+19.9pp
Average	71.7%	58.9%	+12.7pp

Mathematical layers help most where heuristic methods struggle — the harder the conversation, the bigger the improvement.

Ablation (What Each Component Contributes)

Removed	Impact
Cross-encoder reranking	-30.7pp
Fisher-Rao metric	-10.8pp
All math layers	-7.6pp
BM25 channel	-6.5pp
Sheaf consistency	-1.7pp
Entity graph	-1.0pp

Full ablation details in the Wiki.

EU AI Act Compliance

The EU AI Act (Regulation 2024/1689) takes full effect August 2, 2026. Every AI memory system that sends personal data to cloud LLMs for core operations has a compliance question to answer.

Requirement	Mode A	Mode B	Mode C
Data sovereignty (Art. 10)	Pass	Pass	Requires DPA
Right to erasure (GDPR Art. 17)	Pass	Pass	Pass
Transparency (Art. 13)	Pass	Pass	Pass
No network calls during memory ops	Yes	Yes	No

To the best of our knowledge, no existing agent memory system addresses EU AI Act compliance. Modes A and B pass all checks by architectural design — no personal data leaves the device during any memory operation.

Built-in compliance tools: GDPR Article 15/17 export + complete erasure, tamper-proof SHA-256 audit chain, data provenance tracking, ABAC policy enforcement.

Web Dashboard

slm dashboard    # Opens at http://localhost:8765

<details open> <summary>Dashboard Screenshots (click to collapse)</summary> <img src="docs/screenshots/01-dashboard-main.png" alt="Dashboard" width="600"/> <img src="docs/screenshots/02-knowledge-graph.png" alt="Graph" width="190"/> <img src="docs/screenshots/03-math-health.png" alt="Math" width="190"/> <img src="docs/screenshots/05-trust-dashboard.png" alt="Trust" width="190"/> <img src="docs/screenshots/04-recall-lab.png" alt="Recall" width="190"/> <img src="docs/screenshots/06-settings.png" alt="Settings" width="190"/> <img src="docs/screenshots/07-memories-blurred.png" alt="Memories" width="190"/> </details>

17 tabs: Dashboard, Recall Lab, Knowledge Graph, Memories, Trust Scores, Math Health, Compliance, Learning, IDE Connections, Settings, and more. Runs locally — no data leaves your machine.

Features

Retrieval

4-channel hybrid: Semantic (Fisher-Rao) + BM25 + Entity Graph + Temporal
RRF fusion + cross-encoder reranking
Agentic sufficiency verification (auto-retry on weak results)
Adaptive ranking with LightGBM (learns from usage)

Intelligence

11-step ingestion pipeline (entity resolution, fact extraction, emotional tagging, scene building)
Automatic contradiction detection via sheaf cohomology
Self-organizing memory lifecycle (no hardcoded thresholds)
Behavioral pattern detection and outcome tracking

Trust & Security

Bayesian Beta-distribution trust scoring (per-agent, per-fact)
Trust gates (block low-trust agents from writing/deleting)
ABAC (Attribute-Based Access Control) with DB-persisted policies
Tamper-proof hash-chain audit trail (SHA-256 linked entries)

Infrastructure

17-tab web dashboard with real-time visualization
17+ IDE integrations (Claude, Cursor, Windsurf, VS Code, JetBrains, Zed, etc.)
24 MCP tools + 6 MCP resources
Profile isolation (independent memory spaces)
1400+ tests, MIT license, cross-platform (Mac/Linux/Windows)
CPU-only — no GPU required

CLI Reference

Command	What It Does
`slm remember "..."`	Store a memory
`slm recall "..."`	Search memories
`slm forget "..."`	Delete matching memories
`slm trace "..."`	Recall with per-channel score breakdown
`slm status`	System status
`slm health`	Math layer health (Fisher, Sheaf, Langevin)
`slm mode a/b/c`	Switch operating mode
`slm setup`	Interactive first-time wizard
`slm warmup`	Pre-download embedding model
`slm migrate`	V2 to V3 migration
`slm dashboard`	Launch web dashboard
`slm mcp`	Start MCP server (for IDE integration)
`slm connect`	Configure IDE integrations
`slm profile list/create/switch`	Profile management

Research Papers

V3: Information-Geometric Foundations

SuperLocalMemory V3: Information-Geometric Foundations for Zero-LLM Enterprise Agent Memory Varun Pratap Bhardwaj (2026) arXiv:2603.14588 · Zenodo DOI: 10.5281/zenodo.19038659

V2: Architecture & Engineering

SuperLocalMemory: A Structured Local Memory Architecture for Persistent AI Agent Context Varun Pratap Bhardwaj (2026) arXiv:2603.02240 · Zenodo DOI: 10.5281/zenodo.18709670

Cite This Work

@article{bhardwaj2026slmv3,
  title={Information-Geometric Foundations for Zero-LLM Enterprise Agent Memory},
  author={Bhardwaj, Varun Pratap},
  journal={arXiv preprint arXiv:2603.14588},
  year={2026},
  url={https://arxiv.org/abs/2603.14588}
}

Prerequisites

Requirement	Version	Why
Node.js	14+	npm package manager
Python	3.11+	V3 engine runtime

All Python dependencies install automatically during npm install. If anything fails, the installer shows exact fix commands. BM25 keyword search works even without embeddings — you're never fully blocked.

Component	Size	When
Core libraries (numpy, scipy, networkx)	~50MB	During install
Search engine (sentence-transformers, torch)	~200MB	During install
Embedding model (nomic-embed-text-v1.5, 768d)	~500MB	First use or `slm warmup`

Contributing

See CONTRIBUTING.md for guidelines. Wiki for detailed documentation.

License

MIT License. See LICENSE.

Attribution

Part of Qualixar · Author: Varun Pratap Bhardwaj

Built with mathematical rigor. Not in the race — here to help everyone build better AI memory systems.