Knowledge bases for
the agent era.
Built for the era where your agents write more than your team does. Files, sessions, and Stashes. A company brain agents and humans both write into.
Claude Code
Cursor
Codex
OpenCode
OpenclawThe shape of work is changing
Your agents are about to out-produce your team.
Every Claude, Cursor, or Codex run already generates pages of output — transcripts, plans, scratch tables, half-finished documents, dashboards your agent made on its own. Most of it evaporates the moment the session closes.
Stash is the company brain built for that flow. Sessions stream in as they happen. Files give your team and your agents a real filesystem to write into. Stashes turn any slice of that work into a link you can publish or fork into another workspace.
How it works
Sessions. Files. Stashes.
One workspace, two kinds of writer.
The unstructured stream.
Every agent run flows in automatically — prompts, tool calls, artifacts, plan files. Nothing to remember to save.
The structured layer.
Markdown, HTML, tables, folders. Humans and agents both write here. Agents navigate it as a real filesystem through the CLI and MCP.
The shareable slice.
Bundle pages and sessions into one link. Publish to the world, share with collaborators, or fork an external Stash into your own workspace.
The case for this category
“raw data from a number of sources is collected, then compiled by an LLM into a .md knowledge base, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance it… I think there is room here for an incredible new product instead of a hacky collection of scripts.”
Stash is that product. A company brain humans and agents both write into — not a stack of shell scripts wrapped around a folder of markdown.
Use cases
Built for —
- Engineering live docs→coding-agent plans, ADRs, and design notes that stay current
- Company brain→the shared context every agent and teammate reads from
- Research knowledge base→long-running PKBs with sources, transcripts, and tables
- Ops playbooks→release runbooks and on-call procedures
- Brand voice→editorial guidelines and copy standards agents write to
- Personal knowledge management→notes, drafts, and scratch files for a single operator
Where Stash fits
Built where your current tools stop.
Each does part of the job. None gives humans and agents the same workspace.
A single-user Markdown vault. No real collaboration, and nothing richer than .md.
Stash addsReal-time editing across humans and agents, with HTML, tables, PDFs, and any file type.
Pages for humans. Agents can't browse it like a real filesystem.
Stash addsA virtual filesystem the CLI and MCP expose to agents natively.
Files for humans. No structure agents can reason over.
Stash addsAn agent-readable shell over your files, pages, and sessions.
Editing a doc means a clone, a branch, a PR. Fine for source code, painful for anything else.
Stash addsEdit pages in the browser. Agents read and write them directly.
Built to monitor production agents and improve them. The output is traces and dashboards, not work product.
Stash addsWhere the agent's output is the work product, not telemetry to watch.
Per-agent memory in a black box. Doesn't help the human or agent next to you.
Stash addsA shared workspace humans can read and edit, in real time.
See the memory form
Your team's brain,
actually visible.
Every session, page, and table gets embedded into one space. Stash plots them so you can see how your team's knowledge clusters, and which pages have become hubs the tree leans on.
3D embedding projection. Sessions, pages, and tables projected with PCA. Clusters form around topics — not folders.
Files file tree. Nodes are pages, edges are Stashes. Orange nodes are the hubs your agents keep citing.
Features
A workspace shaped like the tools
agents already use.
- VFS
- The whole workspace mounts as a virtual filesystem an agent can
ls,find, andrgthrough the CLI and MCP server. Pages, sessions, and tables — one addressable tree. - Sessions + files
- Agent transcripts get pushed automatically — every prompt, tool call, and artifact — and live alongside the files, tables, and data your team writes. One workspace, two layers, both first-class.
- Real-time
- Humans and agents edit the same files at the same time. No PR flow, no merge conflicts, no per-agent black-box memory. When an agent writes a page, your teammate sees it appear.
- Agentic search
- Semantic and keyword search across pages, sessions, and tables. Agents query the workspace by meaning, not just filename — and follow links between transcripts and the files that came out of them.
- BYO agent
- Plugins for Claude Code, Cursor, Codex, OpenCode, and Openclaw stream sessions in automatically. Any agent that speaks MCP can read pages, query sessions, and publish Stashes from a terminal.
- Native formats
- Markdown, HTML, CSV, PDF, tables — formats agents already read and write. No proprietary doc format to wrap around, no schema to learn.
- Open source
- MIT licensed, self-hostable on your own Postgres. No vendor lock-in, no opaque memory store. Run the same thing we run.
From the Discover feed
Some Stashes
teams are publishing.
A published Stash is a focused slice of a workspace — sessions, pages, and tables — anyone can open. Fork one into your own workspace and it stays live with the source.
RAG over a million PDFs
End-to-end notes from a month of agentic experiments on long-context retrieval. Includes chunking ablations and the evaluator harness.
Auth patterns · Q2
How we converged on per-tenant rate limits, refresh-token rotation, and the worker-pool pattern after three debugging sessions.
Voice-agent onboarding playbook
Live playbook the design + eng team uses when shipping a new voice flow. Updated weekly by the agents that run the user tests.
Claude vs Opus on long-context
Benchmarks, transcripts, and the table of results from a head-to-head on 100k+ token documents. Forked by 47 workspaces.
Open-source release runbook
The exact Stash we follow every Friday — changelog drafting, blog post, social, the whole flow. Fork and adapt for your team.
Customer support deflection memory
Live customer-support knowledge base our triage agent reads on every ticket. Adds three new pages a day on average.
Agent-native
Designed so an agent can actually use it.
Pages are real Markdown, HTML, CSV, PDF — formats your agent already reads and writes. The whole workspace mounts as a virtual filesystem an agent can ls, find, and rg through the CLI and MCP server.
Plugins for Claude Code, Cursor, Codex, OpenCode, and Openclaw stream every session in automatically — no manual upload, no copy-paste.
Give your agents somewhere
to put their work.
Start free in the managed app, or run the whole thing on your own Postgres. Open source, MIT licensed.
MIT · Self-hostable · No vendor lock-in