Notes from the build.
Short, practical pieces about Hydrate - how it's built, what it's for, and what the benchmark data actually says.
-
There was no standup. The launch shipped anyway.
Two days before the Hydrate beta, the checklist was still a week of serial work and I was the lead developer with no team to ask. So I dogfooded the product. Three hours with Claude Code and ChatGPT produced the plan; an orchestrator Claude Code session ran it for the next fifty hours. Forty-eight commits, mostly while I was asleep.
Read it -
Testing memory features without paying the model
The obvious way to test an LLM feature is to call the model and check what it said. Slow, expensive, and when the test fails you cannot tell if the model is confused or the code is broken. Hydrate has a clean seam in its pipeline so half the testing can happen without a model call. Free stage caught five real bugs. Paid stage cost three cents.
Read it -
Two Models, One Rule: How Solo Developers Cut AI Costs Without Cutting Quality
Sonnet for new decisions. Haiku for established implementation. Hydrate is the bridge. The benchmark data behind the 97% cache hit rate that lets Haiku execute against warm context, and why the realistic solo split is 80% Haiku after your first sprint.
Read it -
What 100 developers on Claude Code actually costs in 2026, and what Hydrate does about it
Anthropic quietly revised their Claude Code cost estimates upward. Average developer: $13/day. 90th percentile: $30/day. For a 100-person team that's $343,200 a year before onboarding, churn, or model choice. Here's the full breakdown with our benchmarked savings numbers.
Read it -
What Hydrate does between sessions
Two features, one loop. The dream system runs offline consolidation between sessions: scanning your fact store, detecting contradictions across the full corpus, producing a clarity report. The inner monologue runs during sessions, tracking goals and uncertainty. Here is what the test results show, and how to use both.
Read it -
Cross-vendor AI memory: Mistral Vibe picked up where Claude Code left off
Alice and Bob built a Go API in Claude Code. Carol joined on Mistral Vibe and started with their full project memory on her first tool call. Then she added a logging convention. Then Alice picked it up and extended it. Full round trip: Claude Code to Vibe to Claude Code, different vendors, one shared fact store.
Read it -
Two Claude Code agents, zero file reads, one codebase built from memory.
We ran real Claude Code sessions inside Docker containers (hooks firing, agents receiving injected context) and asked each agent to work from that context alone. Alice built an entire Go REST API from 13 facts and an empty directory. Bob joined cold, listed every convention without reading a file, and closed with: "this is all from injected hook context."
Read it -
We ran 18 sequential Claude Code sessions across three developers. Team sync cut costs 39%.
Canon propagation is the mechanism. Round 2 sessions - where teammates build on each other's work - showed 51% cost reduction and 44% fewer turns. Carol's lint task went from 12 turns to 3. Here's the full methodology.
Read it -
Memory doesn't make AI agents smarter. It makes cheaper ones viable.
12 benchmarks, 5 scenarios, one finding: memory doesn't improve quality - it shifts which model you need. A 2am P0 incident, 1.84x cheaper, identical result. The 96% cache hit rate is the whole ROI story.
Read it -
Tokenmaxxing isn't productivity - and we have the receipts
TechCrunch reported developers with the largest token budgets ship 2x at 10x cost; AI users see 9.4x more code churn. We measured the same gap and shipped a fix. The Hybrid cell ships 7/7 at $0.20/session - 12% of raw-Opus cost.
Read it -
Hydrate is not just a Claude Code hook
Hooks are the fastest-to-ship integration, but they're one of two paths. The other - MCP - works with every agent that speaks the protocol, and that's the path that actually makes Hydrate portable.
Read it -
Hydrate memory is now on Gemini CLI
We shipped hydrate-mcp today. One Go binary, MCP over stdio, every MCP-capable agent gets Hydrate memory. Proven end-to-end on Gemini CLI.
Read it -
How to use Hydrate
The practical guide - what installing Hydrate actually does, the five patterns that make it pay off, and the one case where you should turn it off.
Read it -
How and when to dehydrate
Dehydrate is a hygiene tool for CLAUDE.md, not a cost amplifier. When to reach for it, when to leave it alone, and the honest benchmark number from civichub.
Read it