Grov - Shared and synchronized AI memory + reasoning across teams

@tade_odunlami Thank you! Appreciate it :)

Context window: Grov tracks token usage. When you hit ~90% capacity, it auto-summarizes your progress, clears the conversation, and re-injects the summary + team memories. So you never lose important context.

Sync: When a task completes, Grov extracts the reasoning and syncs it to your team. When any teammate starts a new Claude session, relevant memories get auto-injected into their context - Claude just "knows" what the team already figured out.

For max people, we currently do not have a “max” on team members, but we recommend a 3-5 person team

@osakasaul Honestly, we're still figuring it out based on feedback exactly like this. If you've got a team of 4, here is my email: stef@grov.dev I'll get you set up for free.

@chilarai Yes! When a memory is synced, we generate embeddings (OpenAI text-embedding-3-small) from the goal, reasoning trace, and decisions. When a teammate starts a session, we do hybrid search (semantic + keyword) to find the top 5 most relevant memories and inject them into their context.

@jaydev13 Thanks so much! Great question.

Grov actually operates at a higher level than code style, it captures reasoning, architectural decisions, and constraints rather than formatting preferences.

So if Dev A discovers "we use OAuth-only, rate limit is 100/min" while debugging auth, that knowledge syncs to the whole team. But code style (tabs vs spaces, semicolons, etc.) stays with your existing tools like ESLint/Prettier configs.

This is intentional, we focus on the "why did we build it this way" knowledge that's usually trapped in one dev's head (or lost when they switch contexts).

@imraju Appreciate it! Great questions,

(1) Versioning: We're actively working on this. Right now memories persist until deleted - we store files_touched and linked_commit but don't auto-invalidate when files change. It's under active development.

(2) Efficiency: We've seen tasks go from 10+ min, 3+ subagents used & ~10 files read by the main model to 1-2 min, 0 subagents launched and 0 to ~3-4 files read when context is available (measured by avoiding re-exploration). Haven't formalized token reduction metrics yet. Would love to hear what you're seeing if you try it.

@benjamin_sicard Great questions!

Without any API key:

- The proxy won't start (it requires an Anthropic key to run)

- Basic memory capture has fallbacks, but core features like drift detection and smart extraction are disabled

- Essentially: you need a key for Grov to work properly

Provider support:

- Currently Anthropic-only — Bedrock and OpenRouter are not supported yet

- We use ANTHROPIC_API_KEY directly with the Anthropic SDK

- Multi-provider support (including Bedrock) is on our roadmap but not implemented

TL;DR: Free plan = your Anthropic key. Pro = we handle the API for you. No Bedrock/OpenRouter support yet, but it's something we're considering!

@john_03 The short version: We run a local proxy that intercepts Claude Code requests, injects team knowledge, and captures what Claude learns.

The longer version:
1. When Claude completes a task, we extract:

- Reasoning trace (the "why" behind decisions)

- Files touched

- Constraints discovered

- Key decisions with rationale

We use Haiku to distill this from the session not just "edited auth.ts" but "Extended token refresh from 5→15min because users were getting logged out"

2.Memories go to Supabase (Postgres + pgvector). Each memory gets an embedding for semantic search. Batched uploads, exponential retry, Row-Level Security per team.

3. On new sessions, we query with hybrid search:

- Semantic similarity (embeddings)

- Lexical matching (keywords)

- File boost (1.2x if you're working on same files)

Top 5 memories get injected into Claude's system prompt before it even sees your question, but only if they pass relevance thresholds (no noise).

The result: Claude answers directly instead of re-exploring. We've measured 10 min → 1.5 min per task, and near-zero "Explore agent" launches.

Bonus: We also extract Q&A pairs from sessions. And when context hits 90% full, we auto-generate a summary and inject it so long sessions never lose context, even after auto-compaction.

You can also check out the code as Grov is open-source :) If you have any other questions you can just email me and I'll get back to you as soon as possible.