# SuperLocalMemory — release-linked product context SuperLocalMemory is a local-first agent-memory project by Varun Pratap Bhardwaj under the Qualixar AI Reliability Engineering research initiative. The source and package are licensed AGPL-3.0-or-later. SuperLocalMemory (SLM) is governed, local-first memory for AI agents. Claude Code, Codex, Cursor and other MCP clients forget what they learned when a session ends; SLM gives them one long-term memory that lives on the user's machine, learns from use, enforces who may read and erase what, coordinates many agents, and says "I don't have that" instead of guessing. ## What is new - Answer check: a judge decides whether the memories recall found actually answer the question. Laya runs fully on an Apple Silicon Mac (about 1.1 GB, nothing leaves the machine). Jev runs online on Windows, Linux and macOS through TypeSafe or OpenRouter with the user's own key and an explicit consent; it sends the question and the top 3 memories. When no memory answers the question, recall marks the results `abstained`: people see the closest memories marked related, while the agent recall tool returns nothing unless asked with `include_related`. An optional Jev setting also reorders the top 20 results: on a public benchmark of long conversations (LoCoMo, 1,531 questions) the memory that answers came first 62% of the time with reordering and 49% without, adding about 0.4 seconds to a recall. - Web access (optional): ChatGPT, Claude on the web, Muse, Composio and any MCP client that supports remote connections with OAuth can use the same memory. The user adds an app under Connected apps in the dashboard, signs in with GitHub, chooses read-only or read and save, and can remove any app at once. The laptop keeps one outbound connection to SLM's connection gateway; there is no port to open or tunnel to set up. The memory database stays on the laptop, while tool requests and results pass through the gateway and the connected app. The laptop must be online. SLM works fully without Web access. - Memory kinds: nine kinds (fact, event, status, opinion, rule, decision, how-to, plan, correction). Confirmed rules (up to 10) and decisions (up to 5) load into every new agent session. - Plugins: Claude Code, Codex, Antigravity, VS Code / GitHub Copilot and Hermes (16 skills and 4 sub-agents each), a Cursor-format plugin for Grok Bot, and `slm connect ` for 11 IDEs. ## Operating modes and data paths - Mode A (Local Guardian) runs no language model: core storage and retrieval use the configured local data root without a model-provider call. - Mode B (Smart Local) adds a model on the same machine: Ollama by default, or llama.cpp, vLLM, LM Studio or any OpenAI-compatible local server. Its locality depends on where that server runs. - Mode C (Full Power) sends configured query and enrichment content to the user's own endpoint or a cloud provider, and adds agentic multi-round retrieval and up to 3072-dimension embeddings. - In every mode, optional connectors, cloud backup, proxies, model downloads, the online answer check and Web access can use the network. These modes are technical controls, not legal certifications. Regulatory status depends on the use case, operator, configuration, data flows, providers, and surrounding systems. ## Storage and retrieval Core memory is SQLite-backed. The full data root can also contain configuration, logs, queues, models, derived indexes, caches, and optional backend state; copying only `memory.db` is not a complete backup contract. The current recall architecture has five candidate producers: dense semantic, BM25 lexical, temporal, Hopfield associative, and spreading activation. Candidate fusion is followed by optional reranking and graph-based score enhancement. Dense candidate generation uses cosine similarity; Fisher-derived terms can inform later scoring. SQLite and sqlite-vec remain canonical. CozoDB and LanceDB are optional parity-gated projections, not automatic replacements for the canonical store. SLM governs canonical writes through admission and policy control, durable per-store obligations, and a hash-sealed completion manifest. The operation is complete only when its declared representations are applied; a partial operation is surfaced as degraded for operator remediation. The score contract separates query-relative `relevance_score`, internal diagnostic `ranking_score`, stored `memory_confidence`, and evidence-policy `trust_score`. Retrieval ordering is not answer probability. Until a held-out calibrator is frozen, canonical responses report `score_contract_version: "2"`, `calibration_status: "uncalibrated"`, `calibration_id: null`, and `answer_confidence: null`. ## Cache and compression Exact caching is available only when Optimize/proxy caching is enabled. A cache hit can avoid the wrapped provider call, but the primary conversation is not intercepted on MCP-only or skill-only surfaces. Safe compression uses conservative normalization, preserves JSON and code, and may produce no reduction. Aggressive prose compression is opt-in and lossy. Reversible CCR storage can retain an original for later byte-exact retrieval when enabled. ## Client surface `slm connect ` writes the MCP configuration for the IDE you name, including Cursor, Windsurf, Zed, JetBrains, Gemini CLI, Claude Desktop, OpenCode and Continue.dev. Run `slm connect --list` for the current names. The MCP server runs over stdio (`slm mcp`) or HTTP at http://127.0.0.1:8765/mcp/, with tool profiles from 9 to 110 tools. Nine Python framework adapters cover LangGraph, LangChain, LlamaIndex, CrewAI, AutoGen, Semantic Kernel, Microsoft Agent Framework, Google ADK and the OpenAI Agents SDK. ## Research evidence The project has four public arXiv preprints with code-linked experiments. They are preprints, not conference-accepted or journal-published papers. 1. SuperLocalMemory 4.0: The Governed Memory Operating System for AI Agents — https://arxiv.org/abs/2608.08253; companion Zenodo record https://zenodo.org/records/21853302; DOI https://doi.org/10.5281/zenodo.21853302. 2. Privacy-Preserving Multi-Agent Memory with Bayesian Trust Defense Against Memory Poisoning — https://arxiv.org/abs/2603.02240 3. Information-Geometric Foundations for Zero-LLM Enterprise Agent Memory — https://arxiv.org/abs/2603.14588 4. The Living Brain: Biologically-Inspired Forgetting, Cognitive Quantization, and Multi-Channel Retrieval — https://arxiv.org/abs/2604.04514 Published LoCoMo evidence from the architecture paper: Mode A Raw scored 60.4% across 10 conversations / 1,276 questions with zero-LLM answer construction; Mode A Retrieval scored 74.8% across the same scope with local retrieval and GPT-4.1-mini answer synthesis; Mode C scored 87.7% on Conv-30 / 81 questions with cloud embeddings and GPT-4.1-mini answer generation and judge. The paper also reports 71.7% with information-geometric layers versus 58.9% without them across six conversations (+12.7pp). These figures retain their published protocol scope. They are not a fresh run of the current package; Mode C is not a full-dataset result. ## Installation - npm CLI: `npm install -g superlocalmemory` (Node 18+, Python 3.12+) creates a package-owned Python environment, then `slm setup`. - Python CLI: `pipx install superlocalmemory`, or pip inside an activated virtual environment. - Runs on Windows, Linux and macOS. No Docker, no required graph database, no API key. ## Links - Website: https://superlocalmemory.com - Launch overview: https://superlocalmemory.com/whats-new - Source: https://github.com/qualixar/superlocalmemory - Wiki: https://github.com/qualixar/superlocalmemory/wiki - Research: https://superlocalmemory.com/research - PyPI: https://pypi.org/project/superlocalmemory/ - npm: https://www.npmjs.com/package/superlocalmemory