Your AI Forgets Everything. This Free Tool Fixes It (Hindsight, Tested)

Author: Hyperautomation Labs · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Your AI Forgets Everything. This Free Tool Fixes It (Hindsight, Tested) · Channel: Hyperautomation Labs · 06:13 · Watch on YouTube

Topics: AI agents, Prompt and context engineering, AI developer tools, AI models

Summary

The video tests Hindsight, a free open-source memory tool for AI agents, which stores facts from past conversations and makes them available across sessions and tools. It explains how Hindsight retains, recalls, and reflects on information, then demonstrates its use in bakery recommendations and coding tasks. The tests show it can recall changing facts, reason from prior outcomes, and share project rules across coding agents, although the presenter notes caveats around benchmark claims, model choice, and approving Codex hooks. Hindsight may suit people who repeatedly use AI with the same people or projects, but could be excessive for simple automations.

Key points

Hindsight is presented as a free, open-source memory tool for AI agents.
Its three operations are Retain, which extracts and stores facts; Recall, which searches and ranks memories; and Reflect, which reasons across stored information.
The tool separates facts about the world from its own experiences and attaches evidence to its observations.
In the bakery example, Hindsight recalled the current supplier, delivery day, and the reason the previous supplier was replaced.
It used past marketing results to recommend stopping flyers and a Tuesday discount and investing in Instagram videos.
In a coding test, Hindsight shared checkout rules between Claude Code and Codex, and all 10 tested checkout cases passed.
The presenter notes that Claude Code has built-in memory, so Hindsight’s main advantage is sharing memory across multiple coding agents.
Setup options include installing it for coding agents or running a server with Docker and calling Retain, Recall, and Reflect.
The presenter says a cheap model missed a rule, GPT-5 mini found it, and Codex requires approval of its hooks.
Hindsight is aimed at recurring interactions or projects; its own documentation says it may be overkill for simple automation workflows.

Chapters

Start	Chapter	Summary
00:00	The problem of AI forgetting	The presenter introduces Hindsight and illustrates how a fresh AI chat lacks details from earlier conversations.
01:03	How Hindsight works and benchmark caveats	The video explains Retain, Recall, and Reflect, describes how memories are organized, and qualifies the reported benchmark scores.
02:04	Bakery memory and recommendations	Three tests show Hindsight recalling the bakery’s supplier and history, answering by date, and drawing conclusions about marketing.
03:37	Sharing coding rules across agents	A checkout test compares Codex without memory and with Hindsight, followed by a discussion of cross-agent memory and setup.
05:11	Setup, limitations, and intended users	The presenter covers setup requirements, model and hook caveats, who may benefit from Hindsight, and who may not need it.

Quotes

Retain. After a conversation, it pulls out the facts, who, what, and when, and files them in a memory bank. — 01:03

Reflect. It reasons over everything it knows to reach a conclusion that no single memory contains. — 01:03

It did not just remember. It drew a conclusion. — 03:37

One memory shared by every agent in the project. — 04:09

For a simple automation workflow, it may be overkill. — 05:42

Concepts and entities

Hindsight, Vectorize, Claude Code, Codex, LongMemEval, GPT-5 mini, Cursor, Copilot

Full transcript (original language)

00:00 This delivery robot has never seen these buildings. On its first run, it wanders from office to office until it gets lucky. But, it remembers. Five deliveries later, it knows every floor, and it goes straight there. That is Hindsight, a free open-source memory for AI agents with 28,000 stars on GitHub. So, I tested it. I told Claude Code two rules for a cafe. Then, I opened a different AI Codex and gave it a job. Without memory, it guessed 20% wrong.

00:32 With Hindsight, it already knew 15. Here is the problem. Every new AI chat starts from zero. I asked a fresh chat to do one simple job, write the email to reorder flour for a small bakery. It wrote 290 words with a blank where the supplier’s name should go. It had no idea who the supplier was, or that the owner hates long emails. She had said both months ago. Nobody wrote it down. Hindsight writes it down with three

01:03 moves. Retain. After a conversation, it pulls out the facts, who, what, and when, and files them in a memory bank. Recall. When you ask something, it searches four ways at once, by meaning, by exact words, by how people and things connect, and by time. Then, it ranks the results. Reflect. It reasons over everything it knows to reach a conclusion that no single memory contains. And like you, it keeps facts about the

01:33 world apart from its own experiences. In the background, it merges them into observations, beliefs, with the evidence attached. Does it work? On Long Memory Val, a standard test of long-term memory, Hindsight scores 94.6%. Super memory, 85.9. Zip, 71.2. Plain GPT-4, 0 60.2. One caveat, this is Vectorize’s own chart. Their number was reproduced by researchers at Virginia Tech and the

02:04 Washington Post. The other scores are what each company reports. So, I ran my own tests. Test one, meet Maya. She runs a bakery with three staff and over four months, she tells her assistant four things. June, keep emails under five lines and our flower supplier is Flower Co. July, the Tuesday discount flopped but an Instagram video sold us out. August, Flower Co was late three times, so we switched to Mill and Stone with deliveries on Thursdays.

02:35 September, another video doubled sourdough orders and $300 of flyers brought in about two customers. Now, the same request. Without memory, 42 lines and blanks to fill in. With hindsight, five lines. To Mill and Stone, the usual 20 bags this Thursday. Here is the part that simple memory gets wrong. The supplier changed. A plain search finds both suppliers and the AI can pick the wrong one.

03:06 I asked, “Who is our flower supplier?” Mill and Stone in under a tenth of a second and it kept the reason. Flower Co was late three times. What happened in August? It answered by date, not just by matching words. And this is the dashboard. Every fact with who and when, plus the observations it built on its own. Test two is the one that impressed me. I asked, “What marketing should I spend on next month and what should I stop?” I never told it the answer. It worked it

03:37 out. Stop the flyers. $300 for about two customers. Stop the Tuesday discount. Margins dropped. Put that money into Instagram videos of the baking process. The one thing that worked twice. It did not just remember. It drew a conclusion. Test three is for anyone who codes with AI. I opened a small cafe checkout project in Claude Code and told it two rules. Staff discount is 15% but never on alcohol.

04:09 And the till takes no pennies. So, every total rounds to the nearest 5 cents. Then I switched tools, Codex. A different app, a new session. And it had never heard those rules. Same job and a staff checkout. Without memory, Codex did not ask. It guessed. 20% off and no rounding. With hindsight, the first line said from hindsight memory 15% off non-alcohol items and nearest 5 cent rounding. 10 checkout cases, all passing. One memory shared by every agent in the

04:41 project. To be fair Claude Code now has its own built-in memory. And in my control run, it remembered the rules, too. So, if Claude Code is the only tool you use, you may not need this. The win is one shared memory across Claude Code Codex, Cursor, Copilot, and more than a dozen other coding agents. It also learns from your Git history. Setting it up. If you code with agents, it is one command in your terminal. NPX vectorize io {slash} hindsight

05:11 coding agents install then Claude Code or install all. To add memory to your own app, start the server with one Docker command. Then make three calls. Retain, recall, and reflect. It works with more than 25 model providers including local models and even your existing Claude or ChatGPT subscription with no API key. Two things I learned the hard way. The cheapest model I tried missed the rule completely.

05:42 GPT-5 mini found it. And Codex will not run the memory until you approve its hooks. So, approve them the first time it asks. So, who is this for? Anyone whose AI talks to the same people or works on the same project again and again. Support, sales, a personal assistant, or a team of coding agents. Who is it not for? Their own docs say it. For a simple automation workflow, it may be overkill. I put everything into one PDF.

06:13 The three ways to set it up, the exact commands, my three tests with the real output, and the two mistakes to avoid. Comment hindsight and my bot sends it to you. If this was useful, hit join to become a member here on YouTube or subscribe on Facebook. New to Cloud Code? My beginner guides are in the description. See you in the next one.