qualixar/superlocalmemory

Why SuperLocalMemory?

Every hosted AI memory platform — Mem0 Cloud, Zep Cloud, Letta Cloud, EverMemOS Cloud — sends your data to cloud LLMs by default. Their self-hosted variants exist (Mem0 OpenMemory, Letta self-hosted, Graphiti) but require Docker + a separate graph DB or Ollama config, and most still default to OpenAI until you flip env vars. After August 2, 2026, any of those cloud paths becomes 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 ships local-first out of the box — no Docker, no graph DB, no API keys — on CPU.

The numbers (evaluated on LoCoMo, the standard long-conversation memory benchmark). Published numbers as of April 2026:

How to read this table. Scores from different papers use different LoCoMo splits, judge models, and prompt variants. We do NOT claim these numbers are apples-to-apples across rows. The rows we re-ran in-house are marked "In-house"; cited rows link to the vendor's public source and date. Mode A is the only zero-LLM configuration in the list, so the comparison that is apples-to-apples is Mode A 74.8% vs Mem0 zero-retrieval-LLM 64.2% (+10.6pp). Mem0's 91.6% and EverMemOS's 93.05% use cloud LLMs; Mode C uses a local LLM (Ollama). BEAM-10M, the emerging successor benchmark, will be added in a future release.

What Mode A is: CPU-only, SQLite-only, zero-LLM retrieval pipeline on published LoCoMo questions. To the best of our knowledge it is the only publicly-released local-first memory that clears Mem0's zero-LLM baseline on this benchmark. If another fully-local system hits similar numbers, please open an issue so we can update the table.

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.

V3.3 gives your memory a lifecycle. Memories strengthen when used, fade when neglected, compress when idle, and consolidate into reusable patterns — all automatically, all locally. Your agent gets smarter the longer it runs.

Features at a Glance

• Adaptive Memory Lifecycle — memories naturally strengthen with use and fade when neglected. No manual cleanup, no hardcoded TTLs.

• Smart Compression — embedding precision adapts to memory importance. Low-priority memories compress up to 32x. High-value memories stay full-resolution.

• Cognitive Consolidation — the system automatically extracts patterns from clusters of related memories. One decision referenced 50 times becomes one reusable insight.

• Pattern Learning — auto-learned soft prompts injected into your agent's context at session start. The system teaches itself what matters to you.

• Hopfield Retrieval (6th Channel) — vague or partial queries now complete themselves. Ask half a question, get the whole answer.

• Process Health — orphaned SLM processes detected and cleaned automatically. No more zombie workers eating RAM.

New CLI Commands

# Run a memory lifecycle review — strengthens active memories, archives neglected ones
slm decay

# Run smart compression — adapts embedding precision to memory importance
slm quantize

# Extract reusable patterns from memory clusters
slm consolidate --cognitive

# View auto-learned patterns that get injected into agent context
slm soft-prompts

# Clean up orphaned SLM processes
slm reap

New MCP Tools

Mode A/B Memory Improvements

Embedding migration happens automatically when you switch modes — no manual steps needed.

Dashboard

Three new tabs: Memory Lifecycle (retention curves, decay stats), Compression (storage savings, precision distribution), and Patterns (auto-learned soft prompts, consolidation history). Seven new API endpoints power the new views.

Enable V3.3 Features

All new features default OFF. Zero breaking changes. Opt in when ready:

# Turn on adaptive memory lifecycle
slm config set lifecycle.enabled true

# Turn on smart compression
slm config set quantization.enabled true

# Turn on cognitive consolidation
slm config set consolidation.cognitive.enabled true

# Turn on pattern learning (soft prompts)
slm config set soft_prompts.enabled true

# Turn on Hopfield retrieval (6th channel)
slm config set retrieval.hopfield.enabled true

# Or enable everything at once
slm config set v33_features.all true

Fully backward compatible. All existing MCP tools, CLI commands, and configs work unchanged. New tables are created automatically on first run. No migration needed.

100x faster recall (<10ms at 10K facts), automatic memory surfacing, associative retrieval (5th channel), temporal intelligence with bi-temporal validity, sleep-time consolidation, and core memory blocks. All features default OFF, zero breaking changes.

Enable with slm config set v32_features.all true. See the V3.2 Overview wiki page for details.

Quick Start

Install via npm (recommended)

npm install -g superlocalmemory
slm setup     # Choose mode (A/B/C)
slm doctor    # Verify everything is working
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"]
    }
  }
}

33 MCP tools by default (+42 optional behind SLM_MCP_ALL_TOOLS=1) + 7 resources. Works with any MCP-compatible client — we ship templated configs for Claude Code, Cursor, Windsurf, VS Code Copilot, Continue, Cody, ChatGPT Desktop, Gemini CLI, JetBrains, Zed, and Antigravity (15 IDE configs in ide/configs/). V3.3: Adaptive lifecycle, smart compression, and pattern learning.

Dual Interface: MCP + CLI

SLM works everywhere — from IDEs to CI pipelines to Docker containers. Both the MCP server and the agent-native CLI are first-class, so the same backend serves IDE-side integrations and scripted automations.

Agent-native JSON output on every command:

# Human-readable (default)
slm recall "database schema"
#   1. [0.87] Database uses PostgreSQL 16 on port 5432...

# Agent-native JSON
slm recall "database schema" --json
# {"success": true, "command": "recall", "version": "3.0.22", "data": {"results": [...]}}

All --json responses follow a consistent envelope with success, command, version, data, and next_actions for agent guidance.

Smart-hook architecture (v3.4.43)

SLM ships a small set of Claude Code hooks that fire memory operations only when there's a real signal — not on a timer, not on every keystroke. The hooks are perf-budgeted (<10ms p99 for the hot path) and fail-open (any crash → silent exit, never blocks your prompt). Install them with one command:

slm hooks install      # wires hooks into ~/.claude/settings.json
slm hooks status       # shows what's installed
slm hooks remove       # cleans up, preserves non-SLM hooks

What "smart" means here: the hooks don't interrupt you on a schedule. They watch for specific events that indicate memory work would add value — a topic pivot, a web call about to fire, a re-asked question, a file edit. Otherwise they stay out of your way.

Observability for the new hooks: topic_shift writes one TSV line per decision to ~/.superlocalmemory/logs/topic-shift.log (timestamp | session_hash | current_words_count | window_depth | max_overlap | fired | prompt_preview). Disable with SLM_TOPIC_SHIFT_LOG=0.

Upgrading from v3.4.42 or older: Run slm hooks install once after upgrade to pull in the new wiring. slm hooks status will flag the version mismatch. Merge is idempotent — safe to run twice.

Three Operating Modes

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

Mode A is, to the best of our knowledge, the only publicly-released agent memory that runs with zero cloud calls while clearing Mem0's published LoCoMo score. All data stays on your device. No API keys. No GPU. Runs on 2 vCPUs + 4GB RAM. If another fully-local system hits similar numbers, please open an issue — we'll update this line.

Architecture

Query  ──►  Strategy Classifier  ──►  5 Parallel Channels:
                                       ├── Semantic (Fisher-Rao geodesic distance)
                                       ├── BM25 (keyword matching)
                                       ├── Entity Graph (spreading activation, 3 hops)
                                       ├── Temporal (date-aware retrieval)
                                       └── Hopfield (partial-query completion / associative recall)
                                                    │
                                       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:

1. 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. To the best of our knowledge, the first public application of information geometry specifically to agent memory retrieval — if prior work exists please open an issue so we can credit it.

2. Sheaf Cohomology for Consistency — Algebraic topology detects contradictions by computing coboundary norms on the knowledge graph. We are not aware of a prior production agent-memory system that computes sheaf-cohomology coboundary norms this way; corrections welcome.

3. 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)

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)

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

Ablation (What Each Component Contributes)

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.

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.

Multilingual Embedding Support

v3.4.24+: Plug in any OpenAI-compatible embedding endpoint — Ollama, vLLM, LiteLLM, or self-hosted models like bge-m3, multilingual-e5, Qwen3-Embedding. Configure from the dashboard (Settings > Step 3) or config.json. SLM's math layer (Fisher-Rao, Sheaf, Langevin) is language-agnostic — swap the embedding model and all 30+ languages work at full retrieval quality. No cloud dependency. No code changes. Your data, your language, your model.

Web Dashboard

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

v3.4.4 "Neural Glass": 17-tab sidebar dashboard with light + dark theme. Knowledge Graph (Sigma.js WebGL, community detection), Health Monitor, Entity Explorer (1,300+ entities), Mesh Peers (P2P agent communication), Ingestion Status (Gmail/Calendar/Transcript management), Privacy blur mode. Always-on daemon with auto-start. 8 mesh MCP tools built-in. Cross-platform: macOS + Windows + Linux. All data stays local.

Living Brain Evolution visibility: v3.4.21 ships the reward model, shadow test + online retrain, and evolution cost log via the REST API and slm status --json; the dedicated dashboard tiles are deferred to the next cycle. See docs/DASHBOARD-COVERAGE.md for endpoints and workarounds.

Every recall generates learning signals. Over time, the system adapts to your patterns — from baseline (0-19 signals) → rule-based (20+) → ML model (200+, LightGBM trained on YOUR usage). Zero LLM tokens spent. Four mathematical signals computed locally: co-retrieval, confidence lifecycle, channel performance, and entropy gap.

Auto-capture hooks: slm hooks install + slm observe + slm session-context. MCP tools: session_init, observe, report_feedback.

No competitor learns at zero token cost.

Multi-Machine Mesh Coordination (New in v3.4.48)

Run SLM on multiple machines (M4 + M5) and have your agents coordinate as one team without any disruption.

Setup

M4 (broker):

export SLM_MESH_HOST=192.168.1.100
export SLM_MESH_SHARED_SECRET=my-secret-key
slm init  # Starts SLM at http://192.168.1.100:8765

M5 (client):

export SLM_MESH_PEER_URL=http://192.168.1.100:8765
export SLM_MESH_SHARED_SECRET=my-secret-key
slm init  # Syncs M4's agents every 30s, proxies messages to M4

How It Works

• HTTP-based sync — M5 queries M4's /mesh/peers endpoint every 30 seconds

• Message proxying — When M5's agent sends a message to an M4 agent, it's routed automatically

• mDNS discovery (optional) — M5 can auto-discover M4 on the LAN via _slm-mesh._tcp (enable with SLM_MESH_DISCOVERY=on, default)

• Graceful fallback — Network errors logged but don't crash; offline agents queue messages for delivery

• Shared secret — SLM_MESH_SHARED_SECRET gates remote peer discovery (required for remote mode)

Environment Variables

Dependencies

• zeroconf>=0.140 (new in v3.4.48, optional, pure Python, auto-installed)

• httpx==0.28.1 (already in core deps)

• No Docker. No external broker. Works on WiFi + LAN.

MCP Tools

All 8 mesh tools work seamlessly across machines:

Features

Retrieval

• 5-channel hybrid: Semantic (Fisher-Rao) + BM25 + Entity Graph + Temporal + Hopfield (associative / partial-query completion)

• RRF fusion + cross-encoder reranking

• Agentic sufficiency verification (auto-retry on weak results)

• Adaptive ranking with LightGBM (learns from usage)

• Hopfield completion for vague/partial queries

Intelligence

• 11-step ingestion pipeline (entity resolution, fact extraction, emotional tagging, scene building)

• Automatic contradiction detection via sheaf cohomology

• Adaptive memory lifecycle — memories strengthen with use, fade when neglected

• Smart compression — embedding precision adapts to memory importance (up to 32x savings)

• Cognitive consolidation — automatic pattern extraction from related memories

• Auto-learned soft prompts injected into agent context

• Behavioral pattern detection and outcome tracking

Skill Evolution

• Per-skill performance tracking — tracks which skills succeed and fail across sessions (zero-LLM, always on)

• Evolution engine — 3-trigger system with blind verification. Off by default — enable via slm config set evolution.enabled true

• MCP tools — evolve_skill, skill_health, skill_lineage for programmatic access

• Lineage DAG — visual evolution history in the dashboard

• CLI config — slm config get/set for all evolution settings

• Post-session triggers — automatic analysis on session end via Stop hook

• ECC integration — optional enhanced observations via slm ingest --source ecc

Tiered Storage & Scaling

• 4-tier lifecycle — active, warm, cold, archived with automatic promotion/demotion

• Deep recall — archived facts searchable at reduced weight

• Graph pruning — automatic cleanup of orphan edges, self-loops, duplicates

• Fact consolidation — clusters related facts into consolidated summaries

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.)

• 33 default MCP tools (+42 optional via SLM_MCP_ALL_TOOLS=1) + 7 MCP resources

• Profile isolation (independent memory spaces)

• 2,900+ tests, AGPL v3, cross-platform (Mac/Linux/Windows)

• CPU-only — no GPU required

• Automatic orphaned process cleanup

CLI Reference

Research Papers

SuperLocalMemory is backed by three published research papers (arXiv preprints + Zenodo DOIs) covering trust, information geometry, and cognitive memory architecture. These are preprints — not conference-accepted or journal-published yet.

Paper 3: The Living Brain (V3.3)

SuperLocalMemory V3.3: The Living Brain — Biologically-Inspired Forgetting, Cognitive Quantization, and Multi-Channel Retrieval for Zero-LLM Agent Memory Systems Varun Pratap Bhardwaj (2026) arXiv:2604.04514 · Zenodo DOI: 10.5281/zenodo.19435120

Paper 2: Information-Geometric Foundations (V3)

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

Paper 1: Trust & Behavioral Foundations (V2)

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{bhardwaj2026slmv33,
  title={SuperLocalMemory V3.3: The Living Brain — Biologically-Inspired
         Forgetting, Cognitive Quantization, and Multi-Channel Retrieval
         for Zero-LLM Agent Memory Systems},
  author={Bhardwaj, Varun Pratap},
  journal={arXiv preprint arXiv:2604.04514},
  year={2026},
  url={https://arxiv.org/abs/2604.04514}
}

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

@article{bhardwaj2026slm,
  title={A Structured Local Memory Architecture for Persistent AI Agent Context},
  author={Bhardwaj, Varun Pratap},
  journal={arXiv preprint arXiv:2603.02240},
  year={2026}
}

Prerequisites

All Python dependencies install automatically during npm install — core math, dashboard server, learning engine, and performance optimizations. If anything fails, the installer shows exact fix commands. Run slm doctor after install to verify everything works. BM25 keyword search works even without embeddings — you're never fully blocked.

Contributing

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

License

GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE.

For commercial licensing (closed-source, proprietary, or hosted use), see COMMERCIAL-LICENSE.md or contact varun.pratap.bhardwaj@gmail.com.

Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar.

Attribution

Part of Qualixar · Author: Varun Pratap Bhardwaj

Acknowledgments

• Everything Claude Code (ECC) — SLM's skill observation patterns were inspired by ECC's continuous learning architecture. SLM supports direct ingestion of ECC observations via slm ingest --source ecc, giving ECC users richer skill performance tracking. We recommend ECC for Claude Code users who want the deepest learning experience alongside SLM.

• HKUDS/OpenSpace — The skill evolution research in SLM draws from the EvoSkills co-evolutionary verification concepts (arXiv:2604.01687). We adopted their 3-trigger evolution system and anti-loop guard patterns.

⭐ Support This Project

If this project solves a real problem for you, please star the repo — it helps other developers discover Qualixar and signals that the AI agent reliability community is growing. Every star matters.

Part of the Qualixar AI Agent Reliability Platform

Qualixar is building the open-source infrastructure for AI agent reliability engineering. Seven products, seven research papers (published as arXiv preprints + Zenodo archives), one coherent platform. Each tool solves one reliability pillar:

Zero cloud dependency. Local-first. EU AI Act compliant.

Start here → qualixar.com · All papers on Qualixar HuggingFace