Learn · Playground

Watch memory in motion.

Type your own turn and query — watch the same input flow through write filtering, extraction, storage, retrieval, fusion, and context assembly.

Your inputs — all demos respond to these

Pre-filter scoreevent
0.72 PASS

The read pipeline retrieves from the store built by the write turn above.

01 · End-to-end

One turn, the full pipeline

Step your turn through all eight stages. Each stage computes from the text above — change the input, reset, and step again.

Full pipeline · stepwise
Pick a sample above or type your own. Step forward through write → store → query → retrieve → context.
Stored memory
eventconf 0.69~21 tokens

I just moved to Berlin from Munich and joined Volkswagen on the platform team.

{
  "type": "event",
  "content": "I just moved to Berlin from Munich and joined Volkswagen on the platform team.",
  "confidence": 0.69,
  "valid_from": "2026-05-10",
  "freshness": "1.00",
  "access_count": 0,
  "token_estimate": 21
}

↓ this memory is now in the retrieval pool ↓

02 · Write side

The funnel

100 turns through the seven-stage write pipeline. Watch where the rejections happen. Detail page.

Write pipeline funnel
03 · Read side

Five retrievers, fused

Five ranked lists merge into one via RRF. Your stored memory (marked yours) appears in the retrieval results. Detail page.

RRF fusion
04 · Context assembly

Token budgeting

Retrieved memories must fit a fixed budget alongside system prompt and recent turns. Detail page.

Token budget

Want the detail?

Every stage shown here has a full Learn page with the math, code, and interactive demo.

Back to the Learn hub →

Updates from the lab.

Engineering notes, research drops, occasional product updates. Roughly monthly.