Before
Files and rules multiply; each app remembers its own session.
One memory layer for every agent
Knowit is an open-source MCP server that gives Claude Code, Cursor, and Codex the same persistent project memory — structured, stored in your repo, shared with your team.
Watch it work
One loop: install, capture, recall.
Scroll through the loop your agents run: install once, capture knowledge, recall it anywhere, browse it locally.
01 — Install
One command, every client.
npx knowit install detects Claude Code, Codex, and Cursor, registers the MCP server with each one, and initializes shared memory for the project.
02 — Capture
Agents store what they learn.
While it works, Claude Code calls capture_session_learnings — and the decision lands in Knowit as a structured entry: typed, scoped, queryable.
03 — Recall
Different agent. Same memory.
A brand-new Cursor session calls resolve_context and pulls the exact decision Claude Code stored — before writing a line of code.
04 — Preview
See everything they know.
npx knowit preview opens a local, read-only browser over your memory — filter by type, scope, and tags. No account, nothing leaves your machine.
You explained this last week. And the week before.
Every session starts from zero: the Stripe webhook rule, the deprecated naming pattern, the reason you chose that architecture. Markdown files pile up trying to fix it, and each tool's memory stays locked in that tool. Drag the divider to see the difference.
After
One queryable layer. Every agent reads it the same way.
Set up in three steps.
Run the install wizard.
Run npx knowit install inside your project. The wizard picks your client, scope, and storage, then wires everything up. It can install knowit globally for faster future starts, or skip it and use npx.
npx knowit installYour agent picks up Knowit automatically.
Instruction files and MCP registration are updated for you. Agents call resolve_context before planning and capture_session_learnings after, with no manual wiring needed.
# agent workflow, automatic
resolve_context("add retry logic")
# ... agent works ...
capture_session_learnings()What you get after setup.
Your project now has persistent memory wired into the agent workflow: client registration, project instructions, optional project MCP config, and migrated knowledge from architecture or ADR-style docs.
# after npx knowit install
✓ agent connected to Knowit MCP
✓ project instructions updated automatically
✓ project MCP config added for compatible clients
✓ architecture / PRD / ADR docs moved into Knowit memoryBuilt for teams whose agents share a codebase
Structured knowledge model
Six entry types, four scopes, confidence scores. resolve_context returns the three entries that matter for this task — not a 400-line markdown dump.
Team memory through git
Memory lives in .knowit/knowledge.jsonl — line-delimited JSON with clean diffs. What your agent learned last sprint travels to your teammate's agent through code review.
Every agent, one store
Claude Code, Codex, Cursor, Windsurf, Zed — a first-class MCP server, not a wrapper around one tool.
Your machine, your data
No account, no cloud dependency, nothing leaves your control. SQLite for global scope, JSONL in the repo.
Audit what agents remember
A local read-only browser for the whole knowledge store. Search, filter, read every entry.
Routes to your other docs
When canonical docs live in Notion or another MCP source, Knowit tells the agent where to go next.
The whole tool is free
MIT licensed, local-first, no account. This isn't a trial tier — it's the product.
Open Source
$0
Everything, forever
- All MCP tools and CLI commands
- Unlimited entries, unlimited repos
- Team memory via git or shared SQLite
- Semantic search with your own API key
- Local memory browser
- MIT licensed
Ready in 30 seconds.
One command. The wizard handles everything, including global install if you want it.
npx knowit install