You can now run supermemory locally. Just one command - npx supermemory local Open source binary, available now - git.new/memory

if you've been building on supermemory local, clear some time. it's all happening in the discord first! stay tuned..

love:1 distance:0 Memory Lane with @supermemory

Episode 1 of memory 101: the anatomy of memory.🧠 real memory isn't one thing. it's five: > facts > profiles > relationships > temporality > forgetting "AI memory" products build one of these primitives and stop. Real memory is when all five compose. Next, stay tuned for RAG ↓

So cool. We're on the billboards of SF

kalo saya pribadi gunain integrasi memory dari @supermemory , lumayan menghemat token walaupun sering tiap chat ganti session, agent bakal selalu inget konteks dari session sebelumnya..

if you wanna deep dive into SMFS, read the technical report here: smfs.ai/research/memory-as-a…

@supermemory PROACTIVE agent (Company Brain) ❌ NO MENTION of AGENT ❌ NO LENGTHY PROMPT ✔︎ Understand your own company better than anyone ✔︎ knows what everyone is working on & worked on ✔︎ auto tag people that needs to be involved

supermemory now runs entirely on your own machine! npx supermemory local > full stack runs locally: graph engine, embeddings, fact extraction, user profiles > works with any model (openai, anthropic, gemini, groq) or fully offline via ollama > plugs into claude code, cursor, codex, hermes, langchain, vercel ai sdk

It does and seems to work well with Hermes Agent. Day 2, no major problems. Honcho was different.

paper benchmark, both public: smfs.ai/research/memory-as-a…

Quick explainer on SMFS and why we think agent memory has been measured wrong ↓ everyone optimizes for retrieval speed. but the real cost is how many lookups it takes to answer, often ten tool calls a question, each one burning tokens. SMFS goes after the count. it mounts your supermemory as a real filesystem: → the agent walks it with ls, cat, grep → except grep is semantic, scoped to the folder it's standing in → so it narrows by path instead of re-searching the whole index on xAFS, our open benchmark: → 53% fewer tokens per answer, near-identical accuracy → 6.7 tool calls to an answer, down from 10.2 → at 10k files: 13M tokens the old way vs 2.4M, the gap grows with scale cut the number of retrievals and everything downstream gets cheaper. <speed was never the metric>

Introducing the $100 max plan on @supermemory cloud - 7x limit from pro plan and 10x cheaper than self-hosting same amount of usage - Granola, Gmail and other connectors - Most API features enabled - Best for startups, internal team and heavy personal use!

Start here: github.com/supermemoryai/sup…

→ supermemory.ai/connectors

Your agent only knows what you've uploaded. Docs change, emails land, commits ship, none of it reaches your agent. Connectors fix that. Plug in your sources once. Supermemory syncs the rest. How it works: → You authenticate once. Setup, extraction, and indexing run automatically from there. → Sync is real-time. When something changes on the other end, a webhook reaches Supermemory within seconds. → Only the content that actually changed gets re-processed, never your entire drive. The result is simple: your agents always work from the latest version of your knowledge, with nothing to maintain. Stop importing. Start connecting.

The supermemory chat just got a serious upgrade. • Grok 4.3 alongside GPT, Claude & Gemini • Attach files & images • Thinking mode for deeper reasoning One chat. Backed by @supermemory.

This is something we think and care about a LOT at @supermemory Most people forget that memory is not just about remembering things or storing them (no matter if it's 100k files....) it's also the things on the harness!!! There's tons of people building in memory that almost have psychosis over graph, vectors, whatever stuff. What matters is: - You're giving validated, fresh facts to the agent - Not a static claude.md and this has to be extremely short/compact (we call this a "profile". Not an instructions file, but a living first impression) - The system is great at handling temporality, updated, and contradiction (this is only possible if you have the right data structure and a model to maintain it!) - You're constantly able to verify and reconcile facts (and have an approval process for "inferrred" new things) - @supermemory does this with things like Dynamic dreaming and our custom model harness that's great at long horizon understanding - You also want it to be extremely EXTREMELY configurable per use case. Have dynamic buckets and spaces for different information. Ability to configure what's learnt and to put it into the harness in all different ways - A mass self correcting system or forgetfulness agent built internally to remove all stale info - Great benchmarking and eval platform FOR EVERY USE CASE. Not just longmemeval, my friend. Real world usage doesn't perform well on benchmarks - Sometimes memory is about what you recall (user explicitly asks for things). Sometimes it's about making the agent FEEL good (implicit memory).... balancing the context engineering for this is hard and unsolved imo Given all of this, we have been able to make custom implementations for claude code, openclaw, hermes etc and make the model perform much better and FEEL much better

when friends do epic shit, you gotta pull up

this guy is building @os_vanta, a health OS for people who treat their body like a high-performance machine you plug in everything: ~ whoop, oura, apple watch ~ bloodwork glucose ~ diet, supplements, training it runs on @supermemory, so the system never forgets you, your diet, your symptoms, your preferences, every report you've ever uploaded it reads all of it together and hands you protocols backed by clinical research basically tracking your health like a game you can actually win beta testers have already fixed their sleep, a few are training for hyrox and a marathon on it waitlist is open 👇🏻