A reviewed team memory for Claude Code
Claude Code's built-in memory is a file on one machine: per-repo, per-developer, unreviewed. The fix your teammate's agent discovered on Tuesday evaporates for everyone else by Wednesday. The-Snip turns that memory into a shared, reviewed layer. Agents file what they learn — a gotcha as a short doc, a proven fix as a snippet — into a workspace every teammate's Claude Code searches over MCP. Nothing enters canon until a human approves it, so the memory stays trustworthy instead of accumulating plausible-sounding sludge. The result: your team's agents stop having the same realization independently, and session knowledge outlives the terminal that produced it.
Looking for snippet search and reuse in Claude Code instead? An MCP snippet server for Claude Code
How it works
- Create a workspace as the memory store and generate an API key in Settings → API keys — keys are free; the agents' writes need the Pro & Team plan, $8/user/month.
- Add the MCP server to each repo's .mcp.json with the config below — same key and workspace across repos, so memory is shared, not siloed per project.
- Write the memory convention into CLAUDE.md: before debugging, call search_base for the error or pattern; after solving something novel, file a short doc titled with the exact error message, plus a create_snippet if there's reusable code.
- Let sessions run. Proposals accumulate in the review queue, clearly marked in-review — visible, but not yet canon.
- Triage weekly: approve real lessons, reject noise, and edit titles so search hits them — the error-message-as-title convention makes future lookups land exactly.
- From then on, any teammate's Claude Code recalls the approved fix in one search_base call instead of re-living the debugging session.
{
"mcpServers": {
"the-snip": {
"url": "https://the-snip.com/mcp",
"headers": { "Authorization": "Bearer snip_········_····" }
}
}
}Questions, answered.
How is this different from CLAUDE.md or auto-memory?
Those are per-repo, per-machine files with no review — one developer's context. This is a workspace: shared across repos and teammates, searchable by every agent, and gated so a human approves what the team's agents treat as truth.
Won't agents flood the base with noise?
They'll propose plenty — that's fine, because proposals aren't canon. The review queue is where you keep the signal: approve the genuine lessons, reject the rest in bulk. A tight convention in CLAUDE.md (file only novel, reusable findings) cuts the noise at the source.
What makes a good memory item?
A title someone will actually search — ideally the literal error message — plus two or three sentences of cause and fix, and a snippet if there's runnable code. Small and findable beats long and complete.
Do the agents need seats?
No — agents use workspace API keys; seats are for humans. Reads are free on every plan; agent writes need Pro & Team ($8/user/month), and you can issue and revoke as many keys as you need.
What happens when a memory goes stale?
Anyone — human or agent via update_item — proposes the correction, and it flows through the same review gate. The base's answer stays the current one, which is exactly what an unreviewed memory file can't promise.
Start your base — free.
Free: 25 items, no card. Pro & Team: $8/user/mo — unlimited items, REST API, review workflow, and the hosted MCP server.