Vectorize

Why did we build Hindsight?

It started from us building AI agents for our internal workflows, such as our AI project manager. We quickly realized that our agents needed memory, so we looked around for what was available. We were disappointed with what we found. Most agent memory solutions seemed too basic, basically a wrapper around semantic search, or closed source. But mostly we were disappointed because current solutions didn’t seem to work like how our own, human memories work.

How human memory actually works

Human memories are not static and deterministic. What we remember at any given moment is determined by context and constantly changing memory associations. A human is more likely to remember people and places and is always trying to build a conceptual model of the world. Our memories are time-aware—recency matters, but memories are also anchored in time such that we can usually determine their relative order. And human memory is not just fact-based. We form opinions that evolve over time, shaped by personal dispositions like trust, skepticism, empathy, or rationality.

Designing memory for learning, not storage

With all this in mind, we started building an agent memory system that works like human memory.

From research to Hindsight

Along the way we started talking to folks at Virginia Tech and The Washington Post who saw the same challenges with current memory systems and helped us develop our ideas. The result was Hindsight.

Proving it works

We started using Hindsight with our internal agent solutions and it worked great. We then ran industry-standard benchmarks, like LongMemEval. Hindsight achieved the highest scores ever reported.

Why agent memory is holding back progress

The more we thought about it, the more we realized that current agent memory systems are holding back progress. Many popular agentic tools simply stuff the context window with as much memory as possible and hope the model figures out what’s relevant.

Separation of concerns for agent systems

Providing agents with the most relevant memories for the task at hand—and letting the model focus on execution—produces better results than injecting full histories and user-created files. Separation of concerns is a fundamental engineering principle, and many current tools violate it.

Why we open sourced Hindsight

We decided to make Hindsight open source and freely available to help elevate agentic AI systems by handling memory selection and synthesis, while allowing models to focus on reasoning and task completion.

What’s next

As an open source project, Hindsight is just getting started. We invite you to contribute on GitHub, join the Slack community, and help shape the future of agent memory.