Stash Gives a Team of Coding Agents One Shared Brain

The core problem Stash targets is that every AI coding session generates a rich transcript, and almost all of that context dies on the machine where it was created. This clip from the account @github.signals walks through Stash, an open-source project from Fergana Labs that it found on Hacker News. The pitch is simple: stream every agent transcript and tool call into a shared, searchable layer so a second agent can pick up where the first one stopped instead of relearning the same dead ends.

The Problem Is Lost Context, Not Weak Models

The video opens with a specific failure mode. When you run agents like Claude Code or Cursor, "they generate incredibly valuable session transcripts, but those insights usually stay trapped on a single computer, causing duplicated work and lost context." That framing matters. The bottleneck it describes is not model quality. It is memory. Two engineers running two agents on the same codebase have no way to share what each agent already tried, so both burn tokens rediscovering the same facts and ruling out the same broken approaches.

Stash reframes agent output as a team asset. A transcript is not disposable chat history. It is a record of hypotheses tested and discarded, and that record has value to the next agent that touches the problem.

An Agent-Native Filesystem, Not a Chat Log

The mechanism the video describes is "streaming every transcript and tool call into an agent-native virtual filesystem." The GitHub README frames it the same way and calls Stash "the one place your agents connect to all your data." Instead of dumping logs into a database an agent cannot navigate, Stash exposes memory as something an agent already knows how to traverse: files, folders, and pages.

The frames in the video show the actual README, which lists connectors well beyond raw transcripts: GitHub, Google Drive, Gmail, Slack, Notion, Linear, Jira, Asana, and more. The design goal is one drive where sessions, files, and pages live together, queryable the way an agent queries a filesystem.

Three Capabilities That Compound

The clip names three things agents can do once memory is shared. First, "run semantic searches over past attempts," so an agent can ask what has already been tried rather than starting cold. Second, "edit shared markdown pages in real time," which turns the memory layer into a live workspace instead of a read-only archive. Third, "pick up exactly where another agent left off."

That third capability is the payoff. The value is not any single feature. It is that shared memory compounds across sessions and across teammates. Each solved problem becomes context the next agent inherits for free.

The 50% Claim Is Real, With a Caveat

The voiceover says Stash speeds up engineering tasks "by almost 50%." The README states internal testing "sped up long-running instances of Claude Code by 49%" and cites research by Henry Dowling. That research is the source worth reading directly. Dowling found that without shared transcripts, roughly 48.6% of an agent's work was wasted re-investigating fixes earlier sessions had already ruled out. With access to prior transcripts, agents made about 50% fewer tool calls and cut wasted work by roughly 97%.

So the number is grounded, but read it precisely. The gain shows up on long-running, repeated-context tasks where an agent would otherwise repeat prior work. It is a measure of avoided redundancy, not a universal 49% speedup on every prompt. If your agents rarely revisit the same territory, expect less.

Key Takeaways

• Stash treats AI coding transcripts as a shared team asset instead of throwaway logs stuck on one machine.
• It streams transcripts and tool calls into an agent-native virtual filesystem that agents can navigate like files and pages.
• Agents can semantic-search past attempts, edit shared markdown pages live, and resume another agent's work.
• It connects to GitHub, Drive, Gmail, Slack, Notion, Linear, Jira, Asana, and more, not just agent sessions.
• The headline 49% speedup comes from Henry Dowling's research and applies to long-running tasks with repeated context, where wasted work drops about 97% and tool calls about 50%.
• The project is open source under an MIT license and self-hostable.

Resources

• Fergana-Labs/stash on GitHub — the open-source repo, "shared memory for your team's coding agents," MIT licensed.
• stashai on PyPI — the official CLI, "shared memory for AI coding agents," published by Fergana Labs.
• Techniques to improve coding agent velocity — Henry Dowling's research behind the 49% figure and the wasted-work numbers.
• Original TikTok — the source clip from @github.signals.