Product · memory observability
Every silent decision your agent's memory makes now leaves a receipt
By Rajamohan Jabbala · July 8, 2026
Memory isn't just a store-and-fetch box. Underneath every write and every read, it's constantly making judgment calls: is this worth keeping, is this pattern real or a coincidence, has this stopped being true. Every memory system makes these calls. Almost none of them tell you they happened.
That silence is fine right up until the moment your agent does something you didn't expect, and you have no way to ask why. CLS++ now writes a record for every one of those internal calls, across the whole system, not just the write path you can already see.
The kind of thing that used to be invisible
These decisions were always happening. You just couldn't see them. Now each one leaves a record you can pull up: what happened, why, and to which fact.
Deciding whether something is worth holding onto
Not everything that gets said is worth remembering permanently. Memory now records that judgment call itself, admitted or passed over, instead of just showing you the end result.
Noticing a pattern, then noticing it stopped
Ask the same kind of thing on a regular schedule and memory turns it into a shortcut. Stop asking, and it quietly retires that shortcut rather than keep asserting a routine that isn't happening anymore. Both moments are now on the record, not just the shortcut itself.
Nudging a result because of when you asked
Two facts can be equally relevant on paper; timing can still break the tie. When that nudge actually changes what you see, it's logged once, not silently folded into the ranking.
Why this beats a bigger benchmark number
A recall score tells you the system is usually right. It tells you nothing about the one time it wasn't. An audit trail is the difference between staring at a wrong answer and guessing, versus pulling the exact record of the decision that produced it. For anything that acts on its memory instead of just chatting about it, that difference is the whole ballgame.
The first sense beyond text
Same release, a different kind of first: memory can now take a photo. Send an image instead of typing, and it becomes a fact the same way anything else does, no separate pipeline bolted on the side, no second system to keep in sync. Text stays the only thing every other part of memory has to understand. Audio and video are next on that same path.
Ask your memory why it did something. Get a real answer.
No signup needed to see it work.
I'm onboarding a handful of design partners building agents on a memory layer. If "why did it do that" is a question you're currently answering with a shrug, this is built for exactly that conversation.