Pinned Tweet
Introducing Hindsight Models
Hindsight memory is now 2.5x cheaper and 2x faster, with better answers.
The LLM writes, a small decision model judges. Open weights on Hugging Face, or hosted on the new Hindsight Inference API.
Hindsight memory is now 2.5x cheaper and 2x faster, with better answers.
The LLM writes, a small decision model judges. Open weights on Hugging Face, or hosted on the new Hindsight Inference API.
29
13
137
19,440
You don’t need a big LLM to keep agent memory current. You need a small model that knows what to ignore.
hs-decide-1-flash filters incoming facts so only real changes reach the LLM. glm-5.3-flash went 36 → 84 on consolidation, up to 2.5x faster.
hs-decide-1-flash filters incoming facts so only real changes reach the LLM. glm-5.3-flash went 36 → 84 on consolidation, up to 2.5x faster.
5
2
13
846
Benchmarks are public, run the same way for every model: benchmarks.hindsight.vectorize.io/models
Deep dive in the launch post: hindsight.vectorize.io/blog/2026/11/0…
Deep dive in the launch post: hindsight.vectorize.io/blog/2026/11/0…
0
0
1
63
Vectorize retweeted
Memory work does not always need generated text.
Vectorize introduced Hindsight memory models. The embedding model produces 384-dimensional vectors, with supported reductions to 256, 128 or 64.
#Hindsight #AgentMemory
Vectorize introduced Hindsight memory models. The embedding model produces 384-dimensional vectors, with supported reductions to 256, 128 or 64.
#Hindsight #AgentMemory
1
1
2
50
Vectorize retweeted
A 0.8B decision model nearly matched Gemini 3.8 Flash at checking agent memory, for about 5% of the cost.
In @nicoloboschi's team's test, glm-5.3-flash plus hs-decide-1-flash scored 87.1 for $7.72 per 1,000 cases. Gemini alone scored 89.0 for $147.
hindsight.vectorize.io/blog/2026/11/0…
In @nicoloboschi's team's test, glm-5.3-flash plus hs-decide-1-flash scored 87.1 for $7.72 per 1,000 cases. Gemini alone scored 89.0 for $147.
hindsight.vectorize.io/blog/2026/11/0…
1
1
4
73
Vectorize retweeted
agent memory has a hidden bill: every "is this new? does it matter?" check is a full llm call.
hindsight (mit, 58k stars) just shipped a 0.8b open model that answers those instead. on gemini-3.8-flash: 2.5x cheaper, 1.9x less model time, better scores.
github.com/vectorize-io/h…
hindsight (mit, 58k stars) just shipped a 0.8b open model that answers those instead. on gemini-3.8-flash: 2.5x cheaper, 1.9x less model time, better scores.
github.com/vectorize-io/h…
4
1
1
355
Vectorize retweeted
Remember Hindsight? Back in Sept., I covered how a local 20B model went from 39% → 83.6% on a long-term memory benchmark.
Now Hindsight is making the memory system smarter.
And it's doing it with tiny specialized models.
Hindsight released 3 new open-weight models
👇
🧠 M parameters
🔎 hs-rerank-1 → 114M
💾 hs-embed-1 → 112M
The beauty of Hindsight is that it doesn't ask a large LLM to repeatedly decide if a memory is new or contradictory.
Instead, the 800M model makes those decisions directly. No generated paragraphs. One forward pass.
Hindsight results are impressive
Using GLM-5.3-Flash:
Memory-operation quality
73.8% → 87.1%
Memory consolidation
36% → 84%
Median decision request → 167ms
Across seven tested LLMs:
✅ Six improved in overall quality
💰 Five became cheaper
🔓 All three models have open weights
💻 The 114M reranker runs on CPU
💻 The 112M embedding model runs on CPU
⚙️ The 800M decision model can be self-hosted with vLLM
These are Hindsight's own evaluations, and they're different tests from my September post.
Link in ALT
Now Hindsight is making the memory system smarter.
And it's doing it with tiny specialized models.
Hindsight released 3 new open-weight models
👇
🧠 M parameters
🔎 hs-rerank-1 → 114M
💾 hs-embed-1 → 112M
The beauty of Hindsight is that it doesn't ask a large LLM to repeatedly decide if a memory is new or contradictory.
Instead, the 800M model makes those decisions directly. No generated paragraphs. One forward pass.
Hindsight results are impressive
Using GLM-5.3-Flash:
Memory-operation quality
73.8% → 87.1%
Memory consolidation
36% → 84%
Median decision request → 167ms
Across seven tested LLMs:
✅ Six improved in overall quality
💰 Five became cheaper
🔓 All three models have open weights
💻 The 114M reranker runs on CPU
💻 The 112M embedding model runs on CPU
⚙️ The 800M decision model can be self-hosted with vLLM
These are Hindsight's own evaluations, and they're different tests from my September post.
Link in ALT
A 20B model remembered only 39% of a long-term memory test when its history was stuffed into context.
... but give the same model a new project, Hindsight, and this happens...
🧠 Full context: 39.0%
🧠 Hindsight: 83.6%
Instead of forcing an agent to review its entire history dow every time, Hindsight gives it a separate long-term memory. (super hot GitHub repo)
It can retain what happened, recall relevant memories later, and reflect across those memories to form new conclusions.
And here's why I started digging into it...
You can run Hindsight locally.
It supports:
🦙 Ollama
🖥️ LM Studio
⚙️ llama.cpp
💾 PostgreSQL
Its built-in llama.cpp mode will even auto-download a ~3.5GB Gemma 4 E2B Q4_K_M GGUF.
The benchmark difference gets even crazier on multi-session questions:
😬 Full-context OSS-20B: 21.1%
🤯 Hindsight + OSS-20B: 79.7%
⚠️ These are Hindsight's published LongMemEval results using OSS-20B. They are not results from the ~3.5GB Gemma model.
Hindsight is MIT-licensed.
Link in ALT.
12
10
105
5,061
Improving multilingual ranking was a goal of hs-rerank-1 training.
Turning on @Vectorizeio Hindsight hs-rerank-1 made my agent's recall better in Chinese, English and Japanese. All three, not just English.
Four known memories, twelve checks across three languages: eleven moved up, most of them by dozens of places. In every language, two of the he top 10, and the summary memory is now the first result in all three languages.
One backup memory did not move at all in Japanese. It is in the table too.
That is what multilingual is supposed to mean.
How do you run your memory reranking for multiple languages?
0
1
7
342
See the Exact Prompt Before You Spend a Token.
You write a mission for your memory bank. Something like "focus on billing decisions and who approved them." You save it, and the extraction gets better, or it doesn't.
Now: where did that sentence go?
If you guessed the system r two of the three operations that use it.
TL;DR
- A retain or consolidation mission lands in the user message, not the system prompt. Reflect's lands in the system prompt, where it replaces the agent's role line rather than adding to it.
- POST /banks/{id}/prompts/preview returns the exact messages retain, consolidation or reflect would send. No LLM call, no writes.
- Active blocks concatenate byte-for-byte into the real prompt, and there's a test pinning that against the actual builder.
- A setting you switched off comes back as an inactive block, positioned where it would have landed.
- The endpoint refuses to accept your sample text or a candidate mission, on purpose.
- Shipped in v0.10.0 for self-hosted, and coming soon to Hindsight Cloud.
Know more: linkedin.com/pulse/see-exac…
You write a mission for your memory bank. Something like "focus on billing decisions and who approved them." You save it, and the extraction gets better, or it doesn't.
Now: where did that sentence go?
If you guessed the system r two of the three operations that use it.
TL;DR
- A retain or consolidation mission lands in the user message, not the system prompt. Reflect's lands in the system prompt, where it replaces the agent's role line rather than adding to it.
- POST /banks/{id}/prompts/preview returns the exact messages retain, consolidation or reflect would send. No LLM call, no writes.
- Active blocks concatenate byte-for-byte into the real prompt, and there's a test pinning that against the actual builder.
- A setting you switched off comes back as an inactive block, positioned where it would have landed.
- The endpoint refuses to accept your sample text or a candidate mission, on purpose.
- Shipped in v0.10.0 for self-hosted, and coming soon to Hindsight Cloud.
Know more: linkedin.com/pulse/see-exac…
1
0
7
323
Vectorize retweeted
When we backed Chris Latimer and Chris Bartholomew in 2024, True Partner @puneet324 said they were building an essential piece of the GenAI stack and now their Hindsight Models make that piece faster and cheaper
Congrats to the whole Vectorize team!
Congrats to the whole Vectorize team!
Introducing Hindsight Models
Hindsight memory is now 2.5x cheaper and 2x faster, with better answers.
The LLM writes, a small decision model judges. Open weights on Hugging Face, or hosted on the new Hindsight Inference API.
1
3
7
1,034
Vectorize retweeted
It's amazing what you can accomplish with a team of the smartest, most driven people you've ever met.
✅ SOTA memory performance
✅ Fastest growing agent memory system in the world
✅ The first specialized models purpose-built to deliver SOTA memory faster and cheaper
✅ SOTA memory performance
✅ Fastest growing agent memory system in the world
✅ The first specialized models purpose-built to deliver SOTA memory faster and cheaper
Introducing Hindsight Models
Hindsight memory is now 2.5x cheaper and 2x faster, with better answers.
The LLM writes, a small decision model judges. Open weights on Hugging Face, or hosted on the new Hindsight Inference API.
3
5
20
1,696
Try it in the playground, no GPU needed: inference.hindsight.vectorize.io
0
0
3
221
Self-hosting Hindsight? One login and one environment variable, and Hindsight Inference runs the models for you.
No GPU, $5 of free credit. Or take the open weights from Hugging Face.
No GPU, $5 of free credit. Or take the open weights from Hugging Face.
1
0
6
519
Launch blog: hindsight.vectorize.io/blog/2026/11/0…
Hugging Face: huggingface.co/hindsight-ai
Leaderboards: benchmarks.hindsight.vectorize.io/models
Benchmarks:
benchmarks.hindsight.vectorize.io
Hugging Face: huggingface.co/hindsight-ai
Leaderboards: benchmarks.hindsight.vectorize.io/models
Benchmarks:
benchmarks.hindsight.vectorize.io
0
0
5
492
Vectorize retweeted
@Morrowmake Hermes + Hindsight gets any personal assistant job done. Not that hard to setup and maintain. If, however, you want to emulate a Grok Bot fleet of agents, that's where it becomes a real commitment. The functionality is there, you just need a clear vision of the roster & process.
0
1
2
50
Hindsight: "life changing"
Flipped on hindsight for Hermes and have been playing with advanced delegation and agents last week and it’s been life changing.
I’ve really been able to expand through multiple sessions and keep work moving autonomously for several lanes.
I’m going to try to go headless this week with some options to take over browser sessions if/when needed, feeling more confident expanding given the recent success.
Doesn’t hurt that they make some badass merch too lol
0
0
8
428
Vectorize retweeted
@stretchcloud Hit this exact problem with Claude and Codex not sharing memory. I solved it with Hindsight, and now both remember what we learned.
0
2
2
55
Claude Code Now Builds and Reads Its Own Knowledge Base.
Your coding agent is brilliant and amnesiac. It reasons about a task beautifully, then closes the session and forgets that this repo uses pnpm, that errors are typed Result values, and that you already ruled out a separate nth. Every session relearns the codebase from scratch.
Knowledge Pages, shipped in Hindsight 0.9.0, fix that. They are living documents a memory bank writes about itself, and they rewrite themselves as the bank learns. The part that makes them click for coding: with the Coding Agents integration, your agent builds, maintains, and reads those pages itself. Claude Code does it for you.
Know more: linkedin.com/pulse/claude-c…
Your coding agent is brilliant and amnesiac. It reasons about a task beautifully, then closes the session and forgets that this repo uses pnpm, that errors are typed Result values, and that you already ruled out a separate nth. Every session relearns the codebase from scratch.
Knowledge Pages, shipped in Hindsight 0.9.0, fix that. They are living documents a memory bank writes about itself, and they rewrite themselves as the bank learns. The part that makes them click for coding: with the Coding Agents integration, your agent builds, maintains, and reads those pages itself. Claude Code does it for you.
Know more: linkedin.com/pulse/claude-c…
11
1
16
828