Blog
Notes on memory,
agents & the protocol.
How we think about persistent memory, and how to get the most out of yours.
Curation Is Infrastructure, Judgment Is the Job
Everyone now agrees that agents need a human who cares. But caring is two jobs: deciding what context matters, and judging whether the output is right. Only one of them should stay manual.
Read the post →Transcript Mining Is the New Retrospective
AI session transcripts show where work actually breaks: repeated corrections, missing context, stale instructions, failed checks, and unclear permissions. Teams should mine them before the same mistake repeats.
The Agent-Readable Company
The next operating advantage is not just writing better docs for people. It is making decisions, commitments, incidents, PRs, and transcripts readable by the agents that now help run the work.
Context Is the New CI
Agents are getting faster, but work still depends on state. AI-native teams need a context pipeline that captures, verifies, retrieves, and prunes the facts agents act on.
AI Memory Is Not Enough
Memory helps an AI recall facts. Serious work needs more than recall: typed context, current state, provenance, permissions, and follow-through across every tool and agent.
The chief-of-staff metaphor, and why it matters in a single-prompt world.
A chief of staff used to be reserved for executives. With the right AI tools, everyone can have one: a context layer that keeps receipts, briefs the next room, and makes sure nothing falls between two conversations.
Follow-through is the product.
AI memory is becoming table stakes. The harder product is keeping decisions, commitments, blockers, and source context alive after the chat window closes.
A small evaluation: portable context vs. a fresh model.
The claim is simple: the same model should do better follow-through work when it starts with the right work state. Here is the evaluation design we need before saying that with confidence.
The prompt-compression metric we need before we make compression claims.
Shorter prompts are only useful if they preserve the work state. Here is how we should measure compression, recall, latency, and false carryover before making numeric claims.