Most AI agents forget everything the moment a session ends. Every new conversation starts from zero — no context about the user, no lessons learned, no accumulated knowledge. Hindsight is built to change that. It gives your agents a structured, persistent memory system that learns over time, not just one that remembers raw conversation history.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/vectorize-io/hindsight/llms.txt
Use this file to discover all available pages before exploring further.
Hindsight has achieved state-of-the-art performance on the LongMemEval benchmark — independently reproduced by researchers at Virginia Tech and The Washington Post.
The problem with existing approaches
Simple vector search is not enough for real agent memory. Answering “What did Alice work on last spring?” requires temporal reasoning, not just semantic similarity. Knowing “Alice works at Google” and “Google is in Mountain View” should let an agent answer “Where does Alice work?” even if that exact fact was never stored. And an agent that receives conflicting information over time needs to consolidate that knowledge rather than pile up contradictions. Hindsight addresses these limitations with a memory architecture designed specifically for how agents need to learn and reason.Biomimetic memory hierarchy
Hindsight organizes knowledge into a structured hierarchy inspired by how human memory works:| Type | What it stores | Example |
|---|---|---|
| Mental Model | User-curated summaries for common queries | ”Team communication best practices” |
| Observation | Automatically consolidated knowledge from raw facts | ”User was a React enthusiast but has now switched to Vue” |
| World Fact | Objective facts about the world | ”Alice works at Google as a software engineer” |
| Experience Fact | The agent’s own actions and interactions | ”I recommended Python to Bob last Tuesday” |
TEMPR: multi-strategy retrieval
When you recall memories, Hindsight runs four search strategies in parallel and merges their results using reciprocal rank fusion and cross-encoder reranking:| Strategy | Best for |
|---|---|
| Semantic | Conceptual similarity, paraphrasing |
| Keyword (BM25) | Names, technical terms, exact matches |
| Graph | Related entities, indirect connections |
| Temporal | ”last spring”, “in June”, time ranges |
The three core operations
Every interaction with Hindsight uses one of three operations:- Retain — Store new information. Hindsight extracts facts, entities, and relationships from the content and indexes them for later recall.
- Recall — Search memories. All four retrieval strategies run in parallel and results are reranked by relevance.
- Reflect — Reason over memories. Hindsight uses the bank’s mission, directives, and disposition traits to generate a contextual, grounded response.
Reflect checks memory sources in priority order: Mental Models → Observations → Raw Facts. This means curated knowledge surfaces before raw facts.
Configurable bank identity
Memory banks can be configured with a mission (what the bank knows and cares about), directives (hard rules that must never be violated), and disposition traits (skepticism, literalism, empathy on a 1–5 scale). These settings shape howreflect reasons — not recall — so retrieval stays neutral while generation stays on-brand.
Where to go next
Quickstart
Get Hindsight running and store your first memory in under 60 seconds.
Deploy Hindsight
Docker, Helm/Kubernetes, pip, or embedded in a Python application.
Python SDK
Full API reference for the Python client, including async usage.
Integrations
Connect Hindsight to LangGraph, CrewAI, n8n, and more.