Notes on context, agents, and the layer between.
Context engineering, AI workflows, governance, and lessons we capture while building ContextStream.
Articles
Filter by topic
22 articles
Two superpowers for a coding agent: find the code and remember the plan
A coding agent needs to find the right files and remember what your team already learned. Our new results show why those two skills are much more useful together.
Read the article · 4 min readPortable identity needs portable context
Multi-agent workspaces solve where humans and agents collaborate. They do not automatically solve how approved project understanding follows people across rooms, harnesses, sessions, and tools. Portable identity is only half the continuity problem. The other half is portable context.
5 min readWhy channel history is not enough for cross-agent work
I built ContextStream after spending too many sessions reconstructing the same project for AI assistants. Multi-agent work makes that tax more visible: a room can preserve every message while the next agent still misses the requirement, decision, or constraint that actually matters.
3 min readBefore you refactor, ask the graph
I’ve broken more refactors than I care to admit. Not because the idea was wrong, but because I changed something without knowing what still depended on it, or what was already dead and shipping anyway. Before you refactor, ask the graph: what depends on this, what’s already unused, and where complexity is concentrated.
3 min readThe ContextStream memory audit: 5 questions to ask your agent's recall system
Your agent doesn't fail loudly when memory breaks. It fails politely: re-asks a question you answered in April, re-decides an architecture call you settled weeks ago, ships code that contradicts a convention it "knew" yesterday. Nobody notices until review. Here's a 10-minute audit: five questions, each with a test you can run today.
2 min readInsights are how your AI stops missing the patterns that matter
Some of the most valuable project knowledge isn't a decision yet but rather a pattern someone noticed, a hunch about how the work keeps wanting to grow, an observation that's too soft to be a rule but too useful to lose. Most of that knowledge dies in chat, leaving every new session to rediscover it from scratch. Insights are how ContextStream preserves that middle layer of understanding so your team and your AI agents can keep getting smarter instead of starting cold every time.
10 min readPreferences in ContextStream: The memory event that tells the agent what you prefer.
Most AI friction isn't about what the agent builds, it's about how. Tone, level of detail, planning style: the small corrections you give every session and lose every session. Preferences are how that guidance starts traveling with you instead of resetting every time.
8 min readDecisions in ContextStream: The memory event that tells the agent what's been decided.
AI coding agents reach for "remembers everything" by default, but total recall is exactly what makes them noisy and unreliable. ContextStream was built on the opposite premise: remember the right things and surface them at the right times. The most important "right thing" is a decision, which is a settled, project-level choice your team treats as durable truth. This post explains what counts as a decision, how to capture them with minimal ceremony, and why a hierarchy of decisions, preferences, and insights keeps agents from quietly undoing the choices you have already made.AI coding agents reach for "remembers everything" by default, but total recall is exactly what makes them noisy and unreliable. ContextStream was built on the opposite premise: remember the right things and surface them at the right times. The most important "right thing" is a decision — a settled, project-level choice your team treats as durable truth. This post explains what counts as a decision, how to capture them with minimal ceremony, and why a hierarchy of decisions, preferences, and insights keeps agents from quietly undoing the choices you have already made.
9 min readMy Coding Agent Did the Dumbest Thing. It Was Totally Avoidable.
Last month I watched Opus 4.6 plan a brand new parts database. I already had a parts database. The agent didn't know. That's the tax you pay when your AI has spiky knowledge and no framework filling in the gaps.
6 min readThe brutal cycle that's secretly destroying your velocity with AI
Every developer working with AI in 2026 lives inside the same rhythm. You expand. You contract. Then you expand again. This isn't a flaw in the workflow. This is the workflow. The problem isn't the loop. The problem is what happens the second you close the tab.
3 min readTwo Developers. Claude Code and Codex. One Agent Was About to 'Clean Up' the Fix
Developer One lives in Claude Code. Developer Two lives in Codex. Last week, Developer One's agent refactored the caching layer and made a deliberate call: skip caching for any request flagged as a preview. There was a specific edge case — stale cache was corrupting preview renders. To anyone reading the code now, it looks like an incomplete implementation. This week, Developer Two's agent is about to build a new preview feature. Caching is the obvious move. Without ContextStream, Developer Two's agent confidently introduces the exact bug Developer One spent six hours tracking down. Most teams pay this tax — every new session, every tool switch, every handoff — re-explaining decisions that already happened. Here's what changes when project knowledge finally travels with the work.
4 min readYour Codebase Has Context Your Repo Doesn't Store. Here's Where It Goes.
Most code work slows down before the first edit. The problem is usually not writing the change. It is finding the right code, understanding what depends on it, recovering prior decisions, and keeping follow-up from getting lost. This guide shows how to use ContextStream as a practical layer around a real repo so code work becomes easier to understand, safer to change, and easier to continue later.
7 min readYour AI Agent Isn't Forgetting. It's Reading From a Junk Drawer.
You made the call three weeks ago. Route every Postgres connection through PgBouncer in transaction-pooling mode. You wrote down why. You walked your agent through the trade-offs. Today the agent suggested adding prepared statements. The reflex is to blame the model. Or the prompt. Or whatever framework sits between you and the LLM. The real problem is upstream of all of that — and once you see it, you can't unsee it. Your agent reasons fine. Your knowledge is just a pile.
5 min readWhere Things Go in ContextStream: Memory vs Docs vs Todos vs Plans
New users usually hit the same question quickly: where should this go? If that stays fuzzy, the system gets messy fast. This post gives you the working model for deciding what belongs in memory, what should become a doc, what should be tracked as a todo, and when something is really a plan.
6 min readYour First Hour With ContextStream Is Not About Features
Your first hour with ContextStream should not be about learning every feature. It should be about feeling the core loop once: connect real work, store useful context, retrieve it later, and continue from there. This guide walks through the fastest way to make the product click without getting lost in options.
5 min readWhat ContextStream Is and Why it Matters
AI tools are good at helping in the moment. They are much worse at preserving the state around the work once that moment ends. This post explains what ContextStream is, what problem it actually solves, and why it matters if you want AI-assisted work to accumulate instead of disappearing into chat history.
6 min readNew ContextStream Desktop v0.2 is Now Live — Pushing context engineering forward.
We've nailed cross-session persistence for AI coding tools. Now we're evolving ContextStream into an unmatched platform for context engineering.
1 min readYour Skills Can Now Follow You Across Every AI Tool
ContextStream Skills lets you build a skill once and use it in every AI tool you work in. You can also import skills you've already built or found. No more copying between Cursor rules, CLAUDE.md, and Windsurf configs. As long as the ContextStream MCP is enabled, your skills are there. Team plans let you share skills across your team. Available on all plans, including free.
3 min readThe Missing Layer in Every Multi-Agent Architecture
Most multi-agent designs split work by role. The smarter approach splits by context boundaries. But both assume agents have no memory. Add a shared context layer and the entire calculus changes — you can design by expertise again, and sub-agents stop being a band-aid for context window limits.
6 min readA New Study Says Delete Your CLAUDE.md / AGENTS.md. Here's What Actually Works Instead
A recent paper just measured what CLAUDE.md and AGENTS.md files actually do to coding agents. The results are bad. Developer-written context files improved task success by about 4% on average, while LLM-generated context files decreased success by roughly 3% and both increased cost and latency by over 20%. And this isn't a fringe result it matches what most developers already suspect.
7 min readWhy We Built Notion Integration (And What We Learned)
The hardest part of building AI memory isn't the AI — it's deciding what deserves to be remembered. Here's how we thought through Notion integration and the tradeoffs we made.
7 min readI Got Tired of Re-Explaining My Codebase to AI — So I Built a Memory Layer
After ~18 years of shipping products, I noticed an invisible tax: re-explaining my project to AI assistants every session. Here's how I fixed it.
6 min read