Vectorize — Agent Memory That Learns
Hindsight Sets a New Standard: #1 on the BEAM Benchmark
Agent memory that learns
With Hindsight, your agents remember your users and get better at their jobs over time. Open source, MIT licensed.
The fundamentals
Everything you need from agent memory
- Per-user memory: Every user gets persistent context. Preferences, history, and decisions stay separate.
- Cross-session persistence: Context survives session boundaries. Pick up where you left off, weeks later.
- Fast memory recall: Parallel search returns the most relevant memories in under 100ms.
- Works with any LLM: Model-agnostic memory layer. Swap LLMs without losing what your agent learned.
Agents make mistakes. Yours learn from them.
Every other memory system stores facts and retrieves them. That works until your agent encounters the same problem twice and makes the same mistake both times.
- Learns from mistakes: When a tool call fails or a user corrects your agent, that becomes an experience. Next time, your agent knows what went wrong.
- Detects patterns automatically: The reflection layer synthesizes individual facts into consolidated knowledge. Patterns emerge from data, not from manual tagging.
- Builds judgment over time: Your agent doesn't just get more data. It gets better judgment. Curated mental models guide common situations.
The learning difference, measured
LongMemEval scores across four memory systems. Peer-reviewed benchmark, independently reproduced.
| System | Score |
|---|---|
| Hindsight | 94.6% |
| Supermemory | 85.2% |
| Zep | 71.2% |
| GPT-4o | 60.2% |
Memory that installs itself
Tell Claude Code or Cursor to add Hindsight. It reads the docs, writes the config, and sets up its own memory. One command. No boilerplate.
$ npx add-skill vectorize-io/hindsight --skill hindsight-docs
Installing hindsight-docs skill...
✓ Connected to Hindsight MCP server
✓ Memory tools registered (remember, recall, reflect)
Ready. Your agent now has memory.
Works with any MCP-capable agent. Your agent gets remember, recall, and reflect tools automatically.
Agents are evolving. Memory should too.
Most agent memory systems were built for chatbots. Those days are gone.
Agents that work together, attend meetings, learn from each other, and take on increasingly complex tasks. They need memory that compounds across every agent and every user. Retrieval-only systems won't get you there.
What compound memory looks like
- Shared context across agents: Agent A learns a user preference. Agent B applies it automatically.
- Memory that improves itself: The reflection layer consolidates raw observations into reusable knowledge.
- Per-user across every session: Months of context, instantly recalled. No retrieval lag, no cold starts.
Need help building with agent memory?
We offer hands-on implementation services, team training, and architecture consulting to get your agents learning faster. Whether you're integrating Hindsight into an existing system or starting from scratch, we'll work with you directly.