Curator blog
Product thinking, technical lessons, and release-preparation updates from Curator, evolutionary memory for AI agents. Part of the Code Reality Labs stack.
A health report earns trust by staying quiet when things are fine. Curator's now warns only when work has actually stopped, and it names any captured conversation still waiting to be filed, with the command that files it.
Read post →
A memory benchmark asks whether the right note comes back when you ask for it. That is retrieval. The test that matters day to day is simpler: how often you have to tell your agents the same thing again. That is the measure Curator is held to.
Read post →
An agent searches its memory constantly, and that is good behavior. It should not reshape what the memory holds. In Curator, an agent's search now reads your memory and writes nothing.
Read post →
Encryption at rest has a failure mode nobody advertises: lose the key and a healthy store is noise. Preparing Curator's first public build meant making sure a new machine, a reinstall, or an upgrade never costs you your memory.
Read post →
When Curator opens a project it can bring in defaults you hold elsewhere. That is the most useful thing it does and the easiest to get wrong. It now offers only positions that have held up, that you made, and that bear on the project in front of you.
Read post →
Run enough agent sessions against one memory and the hard part stops being recall. It becomes making sure every save lands. Curator got a round of hardening for exactly that kind of load.
Read post →
As the rest of the stack gets better at producing verified, current context and proving tools correct, a persistent memory layer is worth more, not less. Its job is to keep what was learned so the next session does not start from zero.
Read post →
Long agent sessions get compacted, and the summary usually throws away what you were in the middle of. Curator now puts the working thread back the moment a session compacts, so the agent keeps its footing instead of going blank.
Read post →
Curator's write-time understanding got stronger this sprint. The new model did not swap in on reputation. It had to beat the model already in place, head to head, on a real store, before it shipped.
Read post →
A memory layer stuffed with clever mechanisms sounds better than one that runs only what it can prove. Curator ships the clever ideas disabled when measurement refutes them, treats a revert as data, and keeps live only the parts that earned their place.
Read post →
Retrieval scores are easy to game: return more, and it looks like a hit. Curator grades its recall the way a user actually judges it, so a confident wrong answer scores worse than an honest blank, and a paragraph of near misses does not count as getting it right.
Read post →
Most AI memory asks you to trust that recall is good. Curator benchmarks its own memory on a frozen copy of a real store, with the test and the result that would prove it wrong both written down before the run. Quality you can measure, not quality you assert.
Read post →
Scaling AI engineering breaks on four bottlenecks, and the quiet one is amnesia. Curator removes it with memory that applies what changed since you last touched a project, so agents resume with your current judgment instead of relearning your old mistakes.
Read post →
Most memory only remembers a decision and hopes the agent reads it. Curator puts the recorded decision in front of the agent at the moment of the edit, and in enforce mode hard-blocks the change outright until the decision is honored or superseded.
Read post →
Bulk overnight catch-up can bury the memories that matter. This release ships the guardrail: docs and git history by default, raw transcripts behind an explicit opt-in, and every cleanup reversible because nothing is ever deleted.
Read post →
Teams assume more memory is better, so they import everything, and recall gets worse. Volume buries signal. A memory layer has to rank by durability, ingest selectively, and demote reversibly to keep getting sharper as it grows.
Read post →
A fresh memory knows nothing about a project that already has years of history. Curator can backfill from your git log and docs overnight, so it starts useful instead of empty.
Read post →
Left alone, notes pile up and rot. Curator does the opposite: each night it consolidates what it has learned, so the next day it recalls better instead of noisier.
Read post →
Most memory layers degrade quietly: they'll store noise and rank it next to truth. Curator refuses. No extractable intent means no write; an interpretation outage quarantines the row and tells you, instead of masquerading as a verdict; bad input gets sanitized, not swallowed. Loud failure beats silent rot.
Read post →
Run several AI tools on one machine and they fight over a single graphics card. Curator can now ride a shared local inference backbone instead of loading its own model. By default it uses that backbone whenever one is running and its own model when not, so while the backbone is up, your memory layer and your orchestrator stop competing for the same hardware.
Read post →
Most AI memory hands back whatever is newest or most similar, burying the standards you actually hold. Curator ranks recall by how well-established a memory is, so a durable decision beats this morning's status note.
Read post →
Static memory hoards old facts; RAG retrieves frozen ones. Curator applies what's changed since, and paired with the rest of the stack, your agent finally knows both your code and you.
Read post →
Everyone sells AI memory. Almost all of it remembers what you said and quietly forgets that you changed. Here is the category we are building instead.
Read post →
Curator gives your agent memory over the Model Context Protocol. Memory is one leg of a tripod. Here's what happens when you add the other two.
Read post →
You're not the developer you were a year ago. Your old projects shouldn't pretend you are.
Read post →