Local-first and governed. It says “I don’t have that” instead of guessing.
npm install -g superlocalmemory then slm setup · Windows, Linux and macOS
We're launching on Product Hunt on 15 October
SuperLocalMemory Agent memory that never bluffs: Laya, Jev + Google Gemma Check it out on Product Hunt → 
# Real output from a fresh Mode A install, trimmed
$ slm remember "Staging DB is Postgres 16 on port 5433." $ slm remember "Staging DB moved to Postgres 17 on port 5434." --replaces 0f57a17af3dd46e5 Replaced ✓ 1 fact(s) of 0f57a17af3dd46e5. $ slm recall "which port does staging postgres use" 1. [0.68] Staging DB moved to Postgres 17 on port 5434.
The old memory is kept and the change is undoable.
Works with the tools and bots you already use
Claude Code
Codex
Cursor
VS Code
Copilot
Antigravity
Windsurf
Zed
JetBrains
Gemini CLI
Hermes
ChatGPT
Claude
Composio
Grok Bot
MuseA memory brain, not a bucket: it understands what was saved, keeps it true over time, governs who may see it, and refuses to bluff. Part of Qualixar’s AI Reliability Engineering work.
Every recall ends in one of three tiers: a supported answer, “I don’t have this exactly — here are the closest memories”, or “That isn’t in SuperLocalMemory.” While models load it says it is still warming up and to try again in a few minutes.
How recall decides →Laya, a small on-device model, judges whether the memories found answer the question; it is on by default on Apple Silicon. Jev does the same check online, only after you tick its consent, through TypeSafe or OpenRouter.
Laya and Jev →Screenshots, photos and PDFs become a searchable knowledge base. 
Google’s EmbeddingGemma 2, on your computer: one model reads your words and your pictures, so a description finds the image. OCR runs locally and every answer cites its file and page.
A document index and a document health check, with real hybrid search kept for the moment you need to find something.
See the wiki idea →Memory health describes older memories in the background, only by adding. Anything that moves or deletes waits for your Fix, and Undo puts it back.
Read the guide →Roles, workspaces, save receipts, erasure that reaches every derived copy, a hash-chained audit trail and GDPR export. Built for teams and companies.
See the governance →Every fact knows when it happened and when SLM learned it. Ask what was true last month; corrections stay undoable; unused memories fade.
Time travel →Layered recall gives a one-line glance, the facts, or the full context under a hard token cap, and expand digs deeper only when needed. Cache and reversible compression stop repeat cost.
Every write records which agent made it, with trust scores against memory poisoning, so ten agents can share one memory without poisoning it.
All capabilities →SLM-Mesh gives local agents and web bots messages, locks and shared state, across the internet.
Meet SLM-Mesh →Outcome learning, soft prompts and skill evolution, behind a guarded ranker that rolls back if it slips. Bounded loops and graphs end only when an independent gate passes.
Skill evolution →Four arXiv papers back the layers. The published benchmarks state exactly how they were measured.
Read the research →Every capability, with its guide: the full capability list → · the README
The same governed memory now reaches your bots and chat apps. Your laptop keeps one outbound link; the database never leaves it.
Add it as a Cursor-format plugin on a shared bot computer. No API key, and it runs locally.
Connect through Web access: sign in with GitHub, then choose read-only or read and save.
Connect through Web access and get the same governed recall a local agent gets.
Your laptop must be online. Free daily allowance per connection. SLM works fully without Web access. See every host and plugin →
The agents on your computer and the bots on the web finally share one conversation.
Roles, scopes, GDPR erasure with receipts, retention rules and a hash-chained audit log.
See the governance →Hybrid search across meaning, keywords, the entity graph, time and pictures, a learned ranker with shadow tests and rollback, and time travel.
See how recall works →Laya on your Mac or Jev online judges whether the memories actually answer.
See Laya and Jev →
Every save passes an admission check and gets a receipt. Every read respects roles, workspaces and scopes. Every action lands in a hash-chained audit trail.
Admin, member and viewer per workspace. Company mode attributes every read and write to a signed-in person.
Teams guide →Erasure reaches summaries, caches, indexes and picture mirrors, and comes with a verifiable receipt. GDPR export and erasure (Art. 15, 17, 20).
Compliance guide →A hash-chained audit log, per-profile retention rules and an EU AI Act posture report. Engineering controls that support a compliance program, not a certification.
All the controls →Shared and global only when you choose; cross-profile recall is denied by default.
Scopes guide →Encrypted cloud backup to GitHub or Google Drive. A restore keeps a safety copy and rolls back if it fails.
Restore points →Each feature that sends data says exactly what goes. Pictures, documents, Laya and Mode A never send anything.
Security and privacy →Recall ranks by relevance. The answer check decides whether the top memories actually answer the question.
A small on-device model, about 1.1 GB. Nothing leaves your machine.
Your own TypeSafe or OpenRouter key and an explicit consent. By default it sends the question and the top 3 memories.
An optional switch reorders the top results. Nothing is removed or hidden.
Save a photo, a screenshot or a PDF and find it weeks later: “the slide with the quarterly numbers”.
PNG, JPEG, GIF and WebP. Google’s EmbeddingGemma 2, on your computer, puts text and pictures in one search space, so your words find the image. OCR makes the text inside it searchable. Location data is never read.
Each page is indexed by its picture and its text, and every answer cites its file and page, with a thumbnail.
A document index of every file with its people, projects and dates, and a health check for empty pages, unlinked and duplicate documents and contradicted facts. Hybrid search stays.
The original idea →No generative model ever looks at your files. Pictures live in their own store; erasing a memory erases its picture. Off by default, and it needs 16 GB of memory.
Read the guide →Folders and Obsidian vaults mirror in read-only with slm sources add. Attach a file in ChatGPT, or ask a connected web app for a one-time upload link.
Iteration ends when an independent gate passes, never when the agent says it is done.
slm loop · /slm-loop · MCP toolsLangGraph, LangChain, LlamaIndex, CrewAI, AutoGen and four more share one local store.
See the adapters →See which agent wrote each memory, and every bounded-loop run, in the dashboard.
Roles, sign-in, GDPR export and verified erasure, and an EU AI Act posture report.
See the controls →Isolated workspaces with personal, shared and global memory.
Entities, relations and graph-informed recall.
A ranker that learns from outcomes, with a shadow test and rollback.
Measured skill revisions with lineage, under a budget.
Exact cache and reversible compression, no proxy needed.
Messages, locks and shared state across agents.
One engine behind every surface.
Dashboard for memories, graph, health, mesh and settings.
Gmail, Calendar and transcripts, off until you connect them.
A judge that says “I don’t have that”.
Decisions, rules and how-tos. Standing rules load each session.
ChatGPT, Claude on the web, Muse and Composio.
Ask what was true, and what was known, on any date.
Reusable questions and summaries that cite their memories.
Replace, review and roll back, with restore points.
Change embedding models and repair the store, with undo.
Encrypted cloud backup, and memory served over TLS.
Sixteen skills, four agents, a connector for each IDE, one rulebook.
Published paper evidence (arXiv:2603.14588), not a fresh run of the current package. Mode B has no separate published run.
AGPL v3 · Free local core · Web access has a free daily allowance and is optional
Four public arXiv preprints, not venue-reviewed.
Dense candidates use cosine similarity; Fisher-derived terms refine later scoring.
arXiv:2603.14588 →Lifecycle combines policy, observed use and decay.
arXiv:2604.04514 →Safe mode preserves JSON and code; reversible storage is verified separately.
Product evidence boundary →Plugins for Claude Code, Codex, Antigravity, VS Code / Copilot and Hermes; a Grok Bot plugin; and slm connect for Cursor, Windsurf, Zed, JetBrains and more.
Nine adapters wire each framework’s native memory interface to your local SLM store.
| Framework | Package | What it wires |
|---|---|---|
| LangGraph | langgraph-superlocalmemory | BaseStore long-term memory |
| Semantic Kernel | semantic-kernel-superlocalmemory | VectorStore / RecordStore |
| Microsoft Agent Framework | agent-framework-superlocalmemory | Memory + conversation history providers |
| LangChain | langchain-superlocalmemory | BaseChatMessageHistory + retriever |
| LlamaIndex | llamaindex-superlocalmemory | Chat store |
| CrewAI | crewai-superlocalmemory | ExternalMemory storage backend |
| AutoGen | autogen-superlocalmemory | Memory provider |
| Google ADK | google-adk-superlocalmemory | Memory provider |
| OpenAI Agents | openai-agents-superlocalmemory | Memory provider |
Also in use
Memory they can trust, decisions they can defend, and work that is provably done.
Ends work only when an independent gate passes; multi-step graphs run the same way.
Bounded Loops on GitHub →Installing for a user? Agents can follow SETUP_FOR_AGENTS.md or read llms.txt.
Two commands. Free, open source, local-first.
