Blog

Engineering notes on trustworthy memory for AI agents.

Product · July 17, 2026

We scored our own memory system. Here's the receipt.

A live, clause-by-clause compliance matrix for agent memory. 86.2%, self-assessed against the CLS++ control catalog, with every point we are missing shown in amber. Most memory products hide this. We put it on the wall.

Read

Product · July 16, 2026

Your agent changed its mind. Can you prove why?

CLS++ is durable agent-memory infrastructure — not a consumer chat-sync product. When a belief changes, we archive the old one with lineage so you can audit, debug, and explain what your agents remembered.

Read

Product · July 11, 2026

Store a fact that contradicts an old one, and memory now says so — out loud

Most memory layers silently overwrite a contradicted fact, or keep both and let retrieval guess. Every write into CLS++ was already fact-checked against what memory believed; now the write itself answers back, names the fact it collided with, and keeps both until you settle it.

Read

Product · July 8, 2026

Every silent decision your agent's memory makes now leaves a receipt

Memory quietly decides what's worth keeping, turns a repeated pattern into a shortcut, retires one that stopped repeating, nudges a result because of when you asked. None of that used to be visible. Now every one of those calls writes a record, and memory took its first photo.

Read

Product · July 2, 2026

Actor, Critic, Judge: three models, one memory that can't contradict itself

Most multi-model demos chain models together and call it a day. I built a council instead, grounded in one shared memory instead of three blank slates, so a correction in turn one can't quietly get contradicted in turn three.

Read

Engineering · July 1, 2026

Test your agent's memory in 10 minutes: a reliability benchmark you can run yourself

One pip install and one command reproduce the four failure classes that silently corrupt agent memory — and show a memory layer that survives all four. No signup, runs on your machine.

Read

Engineering · June 26, 2026

The correctness contract: what an agent memory layer should guarantee

Recall is a vanity metric. What decides whether an agent is safe to act on its memory is a correctness contract — a short list of guarantees the layer must never break.

Read

Engineering · June 24, 2026

Four reliability failure modes in agent memory layers — reproduced

Most memory libraries are sold on recall. The failure that actually bites in production is silent loss. Four reproduced failure modes, and how to avoid them.

Read