I've been using Claude Code heavily for a long period of time now. It’s incredible, but I noticed two massive hidden costs that were eating my budget (and my patience).
1. The Context Tax
Every time I started a new session, I had to watch the agent re-explore my codebase. It would read the same auth.ts file it read yesterday, re-analyze the same dependencies, and burn thousands of input tokens just to get back to "baseline."
It felt like hiring a senior engineer who gets amnesia every morning.
2. "Context Drift"
This was the bigger problem. My co-founder would establish an architectural pattern in Session A (e.g., "Always use the Service pattern for database calls"). Two days later, in Session B, my agent would ignore that constraint and write direct SQL queries.
The agent didn't know what my team decided yesterday. This led to regression bugs and "Drift," where the agent slowly deviates from the project's goals.
The Solution: Grov (Team Memory)
I built Grov to give agents a persistent, shared brain.
It acts as a local proxy that sits between your terminal and the LLM. It captures reasoning traces (why a change was made) and syncs them to the cloud.
If developer A explains the Auth system to their agent, developer B's agent automatically knows about it 10 minutes later.
How we built "Anti-Drift"
We didn't just want a vector database dump. We needed active protection against hallucination. We implemented a real-time drift detection system inside the proxy.
Here is the logic flow:
Intercept: The proxy captures every proposed action from Claude (edit, write, bash).
Score: We use a fast, cheap model (Claude 4.5 Haiku) to score the action on a scale of 1-10.
Verify: Haiku checks the action against:
The Original Goal
The Current User Instruction (Takes priority!)
Established Project Constraints
Logical Coherence (Current vs. Historical Reasoning)
Correct: If the score indicates "Drift" (deviation from the goal or constraints), Grov intercepts the request and injects a correction before Claude commits the code.
This effectively gives the agent a "Reviewer" that sits on its shoulder and says, "Hey, remember we decided to use the Service pattern? Don't write that SQL query."
Open Source
I built this because my co-founder and I were tired of repeating ourselves to our AI.
It is fully Open Source (Apache 2.0).
Data: Reasoning traces are captured locally. Team Sync is opt-in.
Repo: github.com/TonyStef/Grov NPM: npm install -g grov
(Star the repo to follow updates!)
We are currently in v0.5 Public Beta. If you are running an engineering team and want to test the "Shared Memory" implementation, I'd love your feedback on the product so far.
Top comments (0)