working on @Vectorizeio Hindsight - SOTA memory for ai agents (100% OSS) hindsight.vectorize.io ASF Member ASF Pulsar PMC ASF BookKeeper committer
Italia
- Tweets 7,942
- Following 1,007
- Followers 1,149
- Likes 8,142
creating a distributed database in rust for petabytes of data wasn't on my bingo card of 2026
we squeezed out as much as possible from postgres and it still works great for mid-sized workloads
but with +70k banks and +50M memories we needed a better solution for us (cloud) and our enterprise customers
39
L'AI italiana deve restare quella che ci fa ridere dietro?
Ho 18 anni e con @yecoai abbiamo detto no.
Sta per uscire Mario-1: un modello proprietario, addestrato da zero, che batte Emma-6.
Ma senza investimenti non potrà essere gratuito per tutti.
yecoai.com/it/investors
2
494
you are the bottleneck if you can't ship this much
I've been doing this for months with great results
I don't know how people say they run 20 agents at once and merge 60 PRs in a day.
That's 4 minutes per PR, assuming 4 hours of actual heads down time. Not only is that not enough time to even glance at the output, it's not enough time to even nail down the requirements first.
That's just pure BS, one way or another.
3
3
354
Congrats Gemini team. Never bet against Google, the company that brought us Transformers and has overwhelming data, engineering talent, and compute.
Everyone has, rightfully, pivoted back to coding. Coding agents are the killer app of AI. We're firmly at an elysium for developers and builders.
Incidentally, getting Muse Wonder and Gemini Argon back-to-back is a reminder why you as an individual, and your company, should stay model agnostic and gateway-maxxx / fx-maxxx.
90
thanks for featuring Giotto
a bunch of other changes I want to make but it completely replaced Claude Design for me
Asking a coding agent to explain an architecture usually gets you a wall of text, or ASCII boxes that fall apart. Giotto gives agents a live browser canvas to draw on over MCP, with automatic layout and no coordinates. Edits are versioned and restorable, and diagrams export as SVG, PNG or animated MP4. Open source from @nicoloboschi, works with Claude Code, Codex, ChatGPT and Claude.ai, local edition MIT declared.
4
145
creating a distributed database in rust for petabytes of data wasn't on my bingo card of 2026
we squeezed out as much as possible from postgres and it still works great for mid-sized workloads
but with +70k banks and +50M memories we needed a better solution for us (cloud) and our enterprise customers
Introducing Hindsight Memlake: agent memory on object storage.
100M memories. Recall under 100 ms p99.
Same Hindsight. Up to 47x faster. Up to 26x cheaper.
1
1
10
363
we built a storage engine for agent memory, and it doesn't have a database.
why: agent memory only grows. an employee's agent writes ~68k memories a year. an autonomous agent running 500 tasks a day writes ~5M. a 10,000-person company with 1,000 agents: ~5.5B a year, and every one needs to be findable in under a second.
we run Hindsight Cloud for thousands of orgs, and at that scale Postgres stops keeping up: on a 210k-memory bank, one recall took 5.3s at p99.
so we built Memlake. everything lives in object storage, in a bucket you control. servers are just caches: lose one, you lose latency, never data. the protocol is model-checked in TLA+.
100M memories on 4 servers, <100ms p99, flat from 1 to 200 recalls/s. for 1,000 autonomous agents, ~$94k/month vs ~$2.46M on Postgres.
and if your banks stay small, Postgres is still great.
open-source Hindsight stays on it, and it's 1.0 today.
Introducing Hindsight Memlake: agent memory on object storage.
100M memories. Recall under 100 ms p99.
Same Hindsight. Up to 47x faster. Up to 26x cheaper.
7
2
16
1,266
more and more harness/agent are shipping better memory into their own products
this is yet another validation from ai labs that memory plays an essential role into agent effectiveness
Hindsight is agent agnostic, you can share knowledge and memories across different harnesses in REAL TIME
Hindsight has a lot of features around consolidation and learning
Hindsight is also cheaper because you can put a very small model as LLM
Hindsight is open source and you can see it and fork it every day
Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.
1
103