← Back to the article A walkthrough · pure MCP-to-MCP
Spinach × MemClaw

Every meeting becomes governed, evolving memory
that an entire agent fleet can trust.

Spinach already turns conversations into structure. MemClaw makes that structure governed, self-correcting, and shared across every agent — so "what was decided" never drifts from "what's true now."

How it's wired

Two MCP servers, one conversation

Both products speak MCP, so the integration is a direct server-to-server handoff inside a single Claude Code session. No webhooks, no Zapier, no glue service to run or maintain.

🥬 Spinach MCP

Source of truth — meetings
  • Summary, decisions, action items
  • Blockers, open questions, owners
  • Chapters, participants, transcript
bridge

🪄 spinach-bridge

Claude Code · type-mapping
  • Maps meeting items → memory types
  • Threads provenance & tags
  • Drives supersession on change
governs

🧠 MemClaw MCP

Governed fleet memory
  • RBAC, audit, tenant isolation
  • Auto-enrichment + PII scan
  • Lifecycle & supersession

What's gone: webhook receiver   middleware pipeline   sync service  —  the two MCP servers talk directly.

The headline

When a decision reverses, memory evolves

A payments squad meets weekly. The second standup overturns a decision from the first. Ask MemClaw "what's true now?" and the reversed decision is simply gone from the answer — not sitting in a pile of contradictory notes.

What's true nowmemclaw_list · status = active
decision
Use Stripe's native reconciliation report
Dana (VP Finance) approved for GA sign-off · weight ↑ 0.95
↳ supersedes the in-house decision
decision
Defer multi-currency to v2 — v1 USD-only
US market launches first
↳ answers the open question
decision
Ship payouts-v2 behind a feature flag
GA target Fri 2026-06-27
blocker ✓
Webhook signature failures — resolved
Root cause: staging clock skew (NTP) · refund QA unblocked
blocker
Payout p99 latency ~4s under burst
Owner: Maya · must clear before GA
stakeholder
Dana requires reconciliation sign-off
Audit non-negotiable
Evolved awaystatus = outdated
decision
Build an in-house reconciliation service
Decided 2026-06-16
✕ superseded by Stripe-native decision
question
Multi-currency in v1, or defer to v2?
Asked 2026-06-16
✕ answered — deferred to v2
These aren't deleted — they're retained, dated, and linked to what replaced them. Full audit trail, just out of the live answer.
What the demo shows

Three beats, all live

1

Real meeting → governed memory

Spinach pulls an actual meeting. The bridge writes it into MemClaw, auto-enriched with title, summary, tags, a PII scan, and full source provenance.

"Pull my most recent meeting and write it into MemClaw with full provenance."
2

A later meeting reverses an earlier one

Week-two standup overturns a decision, resolves a blocker, answers a question. MemClaw supersedes what changed instead of stacking contradictions.

"Reflect the new meeting in MemClaw: supersede what changed, resolve what closed."
3

"What's true now?"

One deterministic query returns the evolved truth. The reversed decision and answered question have dropped out — the fleet reads only current reality.

"Show the continuity pack: what's true vs. what got superseded."
What MemClaw adds on top of Spinach's own recall

Governance, not just retrieval

Governance & RBAC

Tenant isolation, audit, and fleet-scoped permissions enforced on every write — a wrong-fleet write is refused on the spot.

vs. notes anyone can read

Supersession

Changed decisions are marked outdated and linked to what replaced them, so the live answer is always current.

vs. immutable, contradictory minutes

Provenance

Source URI, meeting ID, and item ID thread from transcript → memory → answer. Every claim is traceable.

vs. "where did this come from?"

Auto-enrichment

Every write gets an LLM title, summary, tags, entity graph links, and a PII scan — for free, from one line of content.

vs. manual tagging

Fleet fabric

One team's standup becomes the whole agent fleet's working memory — shared, scoped, and consistent.

vs. siloed per-tool recall

Insights & keystones

Mandatory policies and reflective insights ride alongside memory, steering how every agent acts on it.

vs. tribal knowledge
2
MCP servers · 0 middleware
14
governed memories written live
2 → 0
contradictions in the live answer
100%
provenance retained
1
wrong-team write blocked on the spot

Governance you can see. During the run, a deliberate write to a fleet the credential doesn't own returned FORBIDDEN — fleet-scope policy. Isolation isn't a setting buried in config; it's enforced on every single write.