hindsight
Hindsight is a newer entry from the Vectorize team, organized around three operations: retain (store), recall (retrieve), and reflect (summarize and update mental models). The 'reflect' step is what sets it apart — instead of just storing memories, it periodically rereads them and updates the higher-level model of who you are and what you care about.
- Teams wanting a memory that builds opinions, not just facts
- Long-running assistants where mental models matter
- Developers willing to adopt newer tooling for stronger recall
What you'll do
Hindsight runs as an MCP server via npx. No install step — just point your client at it. Budget 10 minutes the first time.
Before you start
- Node.js 20+
- An MCP-speaking agent client
Step-by-step install
- 011. Add the MCP config
{ "mcpServers": { "hindsight": { "command": "npx", "args": ["-y", "hindsight-mcp"] } } } - 022. Restart the client
Fully quit and relaunch. First run downloads via npx (~20 seconds).
- 033. Use the three operations
Hindsight exposes retain, recall, and reflect as distinct tools. Start with retain: 'Retain that I prefer morning meetings and dislike calendar invites without agendas.' Then recall later: 'Recall my meeting preferences.'
- 044. Trigger a reflect cycle
After accumulating 20+ memories, ask your agent to run the reflect operation. Hindsight re-reads memories and updates the high-level mental model — producing a summary document of who you are.
Tip: The reflect step costs tokens. Don't run it every session; once a week is enough for most users.
Your first 10 minutes
- 01Retain 10 things about you: preferences, priorities, people, projects.
- 02Recall them in a new session. Confirm retrieval.
- 03Run reflect. Read the generated model. It should feel like a shockingly good summary of you.
- 04Hand-edit the model if it's off — hindsight supports direct edits.
- 05Add Cognition CLO on top.
Troubleshooting
Reflect produces a model that feels generic.
Reflect gets better with more source memories. Retain more specific, differentiated facts — not just 'I like AI' but 'I'm betting on MCP specifically because of its portability story'.
hindsight holds the knowledge. Cognition CLO models retention per employee per concept using a Weibull forgetting curve — so you see decay before it becomes a missed SOP or a failed audit.