I Planned a Context Lake for My Coding Agents and Haven't Needed It Yet
One weekend I finished 22 epics using Claude Code on my Veneer project, and I did most of it from my phone over remote control. One of those sessions spun up a themes repo, a repo listing the plugins, and a repo for feature requests and bug reports, along with a few other extras. It knew where my local repos lived. It pushed them to remote.
Here’s what makes that weird. A few months earlier I’d decided my agents needed a whole context system before they could work across projects like this. One place that held the notes, plans and research for every project, so an agent working on one app would know about the apps that talk to it. I wrote about it. I started building it.
I haven’t finished it, and I haven’t needed it yet. I’m not saying I never will. I’m saying it surprised me how far I got without it, and that’s what this post is about.
- The Context Lake I Was Going to Build
- I Just Pointed at Things Instead
- Where It Broke
- What Was Actually Holding It Up
- A Don’t-Lose-It Ledger, Not a Knowledge Base
- What the Context Engineering Crowd Warns About
- The Realistic Limits
- So Am I Building the Lake?
The Context Lake I Was Going to Build
Back in July I wrote I Got Tired of AI Memory Hype, So I Built a Context Lake. The idea was one brain outside the agents that they all read from. I wanted to build it myself instead of picking one of the 50 bazillion options out there because I wanted to know how it worked. I started building it in July, then left it. Anything dated October is me dropping notes in, not code.
Veneer is the plugin-first writing app I rewrote from scratch, and two of its plugins wrap older projects of mine. BookForge is an EPUB job server. GitWrite is a local-first adapter over plain git that stores annotations as empty commits on a feedback branch. I built GitWrite and BookForge about a year and a half ago, when agents wrote worse code than they do now. Getting all three to line up felt like more of a job than “here’s a repo, good luck.”
Then building the lake started taking longer than the work it was supposed to support. So I stopped and did the work.

On October 30 I told an agent in a Veneer session that I needed “some sort of context lake” because my projects were starting to interact. I can’t remember what set that off. The agent logged it on the punch list as “A session in one repo can’t see the others” and parked it as outside Veneer’s scope. It’s still parked.
I Just Pointed at Things Instead
Instead of a system, I pointed. I’d tell an agent to look at this project, or that one, or the Obsidian vault, and let it dig through the maze itself. It looks haphazard. The results held up anyway.
Two pasted paths and a plugin plan
On October 18 I was in a Veneer session and the agent needed to understand the two plugin projects. First it went looking on its own. It searched ~/node, ~/code and ~/dev and came up empty, because both lived in ~/python. So I pasted the paths to git-write/ and book_forge/. It read both repos and wrote Epic 9 plus two plugin audit docs out of what it found.
The brief flipped the plan
This is the cleanest case. Around October 24 and 25, I was in a GitWrite session and sent an Explore subagent over into the Veneer repo. It came back with a ready-made brief for the GitWrite plugin that an earlier Veneer session had left in Veneer’s docs/handoff/ folder. That brief flipped the GitWrite plugin’s whole plan from a remote API to local-first. One agent read another agent’s notes in a different repo, and the plan changed because of it.
Where It Broke
Telling the truth now, because “it worked” covers a lot of mess.
The agent guessed the wrong directories twice, and both times the fix was me handing it the paths. That’s me being the map.
Stale information got through, too. GitWrite and BookForge each had a VENEER_DIRECTION.md note, and both still said “remote API” after the plan had flipped to local-first.

So there’s a fair objection to all of this: maybe pointing works because I’m the index. When the agent can’t find something, I know where it is. I don’t think that’s the whole story, which is the next section, but it’s part of it.
What Was Actually Holding It Up
“Haphazard” undersells it. When I looked back at why pointing worked, most of it came down to things I’d set up for other reasons and stopped thinking about.
Repos that stayed where they landed. I never designed where my repos live. The node/ and python/ folders predate AI. Things stayed put partly because nobody wants to move thousands of node_modules files. The vault structure, on the other hand, was somewhat systematic.
A habit that went away. I used to depend on “copy full path” in VS Code and Obsidian. At some point the agent just knew where things were and I stopped doing it. I have a few minimal Obsidian skills for the CLI and search, but mostly the agent guessed my folder conventions and put things in the right place.
A skill written for a dumb model. My vault-writer skill exists because of OpenClaw running an old DeepSeek model that kept skipping rules. So I hardened the skill until it couldn’t skip them. It syncs to every tool through skillshare, so Claude uses it too, as a side effect. These days I can message OpenClaw on Telegram and have it update the vault from my paper notebook. Context written for the dumbest agent you have benefits every smarter agent after it.
A pointer from the repo to the vault. My living-plan skill lets a repo’s CLAUDE.md name an outside task mirror, which is a note in the vault. GitWrite’s points at the TODO list in its vault note. The rule is that the pointer has to say when to reconcile the two, because a mirror that doesn’t get reconciled is just agent busy work.
Handoff docs written for the other repo’s agent. Veneer’s docs/handoff/ folder holds briefs for both plugins. They’re written for whatever agent opens the other repo next, which is how the GitWrite plan got flipped.

A small API surface on purpose. I planned to build the core app and the plugins at the same time, because each needed things from the other. I modeled the plugin API on Obsidian and VS Code, dumped every plugin idea into one bucket, picked the ones to build first, and had the agent derive only the APIs those plugins needed. Less surface means less for any agent to get wrong.
BMAD inside the repo. Everything above is about getting an agent to the right folder in the right project. Once it’s inside Veneer, BMAD keeps it on track. A fresh session picks up the next story instead of waiting for me to explain where we left off.
None of that is a context system, but all of it is context.
A Don’t-Lose-It Ledger, Not a Knowledge Base
The design principle under all of this took me a while to put into words. New ideas and changes land in Obsidian first, because that’s the convenient place to put them: the vault is on my phone, and dropping in a to-do beats spinning up an agent to talk to the repo just to save a note. So the flow mostly runs from the vault to the project. Updates do flow back, but I don’t sweat that direction, because the repo is the source of truth and the vault just collects.
What the vault tracks is what hasn’t been checked off yet. It’s a ledger so nothing gets lost.
What actually happened is a separate question, and I answer it from the Claude Code session logs. I assume things happened because I see the results, and the logs say the agent did something else. On October 25 I corrected the agent about two different things it had stuck together. One was the GitWrite note being old. The other was that I’d retired kanban boards in my vault for plain notes with to-do lists, a change I’d made across the whole vault. The agent had to be told that one.
What the Context Engineering Crowd Warns About
Context engineering is having a moment, and most of the coverage is rigorous systems and failure modes. The taxonomy that keeps coming up is Drew Breunig’s four ways long contexts fail: poisoning, distraction, confusion and clash. I don’t disagree with any of it. That list fits what I run into.
The clearest one in this story is clash. Two direction notes said “remote API” while the plan said local-first, and an agent reading both would have to pick one. Poisoning is the risk behind it: a stale note that gets read early and then built on.

The fixes Breunig lays out in his follow-up, How to Fix Your Context, include offloading notes somewhere outside the context, pruning, and summarizing a long thread so a fresh one can start, which is basically resetting. They’re mostly what I was already doing without calling it anything. Handoff docs are offloading. BMAD’s story files are resetting. The gap is pruning. Nothing in my setup deletes old material. The stale direction notes got patched because a session happened to trip over them.
The Realistic Limits
When I say the plugins work, here’s what that means. They install, and I use them myself. I run the servers locally. There are no public instances. The services are basic on purpose, there’s no formal end-to-end test, and the only user is me.
I also haven’t measured tokens. My sense is I wasn’t burning a huge number, but that’s a sense, not a number.
So Am I Building the Lake?
Maybe. I’m still considering it. There could be things I can’t see from here. A context lake might get agents to the right place faster. It might use fewer tokens. I don’t know yet.
For me to build it, it would have to work close to perfectly and save real tokens. Or it would have to polish everything I already do. I’m not going to settle that by guessing, so the next step is measuring what just pointing at things actually costs.
Part of what I wanted the lake for was one agent sitting above all my projects. Claude Duet covers a lot of that now. It’s Claude’s take on the personal assistant agent, the same family as OpenClaw or Hermes, meant to connect to all your stuff. It’s built more for project managers and knowledge workers than builders like me, and I don’t feed it much. I don’t have a flood of email or a packed calendar. It has helped with freelance work, like addressing client comments on Google Docs for articles I’d submitted. The part that matters here is that it can see every Claude session I’m running over remote control, so I can orchestrate all of them from one Duet chat. I’m still testing it, and it isn’t exactly what the lake would have been. But one thing that came along with it was a handful of weekly resets. Claude Code gives you a five-hour session allowance and a weekly allowance, and the reset zeroes both, so I’ve been pulling two weeks of work out of a Pro account every week.
For getting the job done, though, I haven’t needed the lake yet. Twenty-two epics in a weekend, mostly from my phone, and the context system I planned for it is still a vault note and a repo with a last commit in October. While I’m writing this post in Veneer, Claude’s working on epic 64 in the background.
