Skip to content
No black box · full traceSee what it remembers

See exactly what your AI remembered — and why.

No black box. For any memory, replay the whole story: the raw input it came from, the fact we pulled out, and why it got recalled for a given question. When something looks off, you can see it — and fix it.

One memory, raw input → recall
  1. Researcher agent transcript: "The client moved the launch to March 14 — they confirmed budget is now $48k, up from $40k."

    source: agent · Researcherreceived: 2026-05-02 14:08
The problem[1 / 6]

Memory you can't see is memory you can't trust.

With most AI memory, you have no idea what's in there, where it came from, or why your assistant said what it said. If it remembers something wrong, you're stuck guessing. We made memory glass-clear.

The trace[2 / 6]

Raw input → fact → recall.

One real memory, walked from the raw input all the way to the answer it shaped — across two agents that share the same memory.

Trace / Replay — how this memory got here
  1. Researcher agent transcript: "The client moved the launch to March 14 — they confirmed budget is now $48k, up from $40k."

    source: agent · Researcherreceived: 2026-05-02 14:08

Click any step to see what happened — including what we chose not to keep, and why.

Provenance[3 / 6]

Where every fact came from.

Every memory carries a receipt: which agent, tool, or document it came from, when, and how many independent sources back it up.

Provenance & corroboration
  • user
  • agent
  • tool / web
  • Project budget = $48,000conf 0.94×2
  • Launch date = March 14, 2026conf 0.90×1
  • Account owner = Dana Leeconf 0.97×3
The read side[4 / 6]

Why it was recalled.

For any question, see the exact memories we considered, the ones we used, the ones we dropped to keep the answer clean — and why each made the cut or didn't.

  • INJECTEDProject budget = $48,000top-ranked, current truth
  • INJECTEDApprover = Dana Leeneeded to action the answer
  • DROPPEDProject budget = $40,000superseded by the current value
  • DROPPEDLaunch date = March 14off-topic for a budget question
  • DROPPEDVendor preference = Acmebelow the token budget after ranking

Token budget 180 / 600 used — the ranking is auditable, line by line.

The hard part[5 / 6]

Misses are visible too.

The scary failure is the silent one — when your AI should have remembered something and didn't. We surface those: a recall that comes back empty or unsure is logged, so a gap becomes something you can see and close.

Transparency → action[6 / 6]

Fix it on the spot.

Spot a bad extraction or a miss? Re-run it over the original raw input with a better model — the source is never thrown away, so memory is always recoverable.

This is the same Trace/Replay that powers our eval harness — explainability isn't a demo, it's the engine.

See what it remembers

Trust what your agents remember.

Every memory traceable back to its source. Nothing mysterious — and nothing thrown away.