SuperLocalMemory Offers Governed, Local-First Memory for AI Agents

Summary

SuperLocalMemory is an open-source, local-first memory control plane designed to give multiple AI agents and teams one governed memory store. It supports Claude Code, Codex, Cursor, Copilot, ChatGPT, Claude on the web, and other tools through MCP or command-line integration. In its core Mode A, remembering and recalling text require no model-provider call, API key, cloud account, or Docker, while optional features can send data out only after the user enables them. The system uses hybrid retrieval across semantic, keyword, entity, temporal, associative, and image channels, and returns one of three outcomes: a supported answer, up to three closest memories labelled as approximate, or an explicit statement that the memory is not present. It also records corrections, supports time-based queries, applies token budgets and layered recall, and treats recalled content as untrusted evidence for prompt-injection defense. Optional picture and document search uses Google’s EmbeddingGemma 2 locally, with local OCR for images and PDF pages; answers cite the source file and page, and the feature requires about 16 GB of memory. On Apple Silicon, the local Laya model can check whether retrieved memories actually answer a question; Jev provides an online alternative only with consent. Team controls include roles, workspaces, scopes, attributed writes, Bayesian trust scores, erasure across derived copies, receipts, retention rules, and a hash-chained audit trail. SLM-Mesh lets permitted local agents and web bots exchange messages, waits, locks, and shared state, while bounded loops require independent completion gates. The project supports macOS Apple Silicon, 64-bit Windows, and 64-bit Linux, and can be installed through npm, pip, an install script, or an agent-guided setup. Its published architecture-paper benchmarks report 60.4% on LoCoMo for raw Mode A retrieval, 74.8% for Mode A with GPT-4.1-mini answers, and 87.7% for one Mode C subset, but the repository cautions that these figures are not fresh runs of the current package and are comparable only under matching settings. The project is licensed under AGPL-3.0-or-later.