Stash – CLI to search over your team's coding agent sessions
Shared memory for coding agents stops teams from re-debugging solved problems.
Team visibility layer for AI agent conversations
Git-linked agent conversation archive, but ClickHouse dependency limits adoption.
Engineering teams using Claude Code, Cursor, or other AI coding agents who want shared visibility into AI-assisted development.
agentsview (local predecessor) · Cursor Sessions · Continue History
I built this because I wanted to understand my own agent usage patterns. I mostly use Claude Code, and agentsview was already doing the hard work of collecting sessions locally.
Then I wanted to sync sessions across my machines so everything was searchable in one place. When I found interesting new approaches with coding agents (superpowers, gsd, higher-order prompting) that I wanted to share with friends and teammates. I ended up creating gists, which is a terrible way to share conversations.
These clicked together into agentlore. A daemon watches agentsview's local database, masks secrets, and syncs conversations to ClickHouse. A web UI lets you browse, search, and filter across all your sessions.
The part I'm most excited about is git linking. Agentlore parses tool calls for git commit and gh pr create invocations, extracting commit SHAs and PR URLs. So you can go from a commit to the full conversation that produced it. Every conversation gets a shareable permalink, so you can drop it in a PR comment or Slack and people can see exactly what happened.
What I'm hoping to see play out:
- Review agents should get more accurate when they have the conversation context, what was intended and what was manually tested, not just the diff. - People can learn from each other. Someone figures out a better way to prompt for refactoring, the whole team can see how they did it. - Context stops disappearing. When you pick up code someone else wrote with an agent, the conversation is right there instead of lost on their machine.
This is early and likely broken in places. Would love feedback.
Demo: https://demo.agentlore.io/sessions/589b7dd5-93ad-4dc0-8e6f-d... (the conversation the built this project. you can see all subagent sessions, but not all commits are captured yet)
Shared memory for coding agents stops teams from re-debugging solved problems.
50k lines shipped in a week, but team knowledge bases are a crowded category.
Path-scoped memory beats project-wide rules files for team context sharing.
Parses local JSONL session files from Copilot, Claude Code, and Codex and surfaces an interactive DAG, per-agent/token metrics, and session replay directly in the activity bar — immediately useful for spotting which agents and skills actually do the work. The Claude cache-token breakdown and timeline replay are clever, concrete features that show the author dug into provider internals; adoption looks small today, but the concept and implementation give real observability where previously there was opacity.
Hook-based capture beats OTEL for Claude Code agent visibility.
npm install for agent skills that copies prompts into Claude, Cursor, and Gemini instantly.