Plumb
v0.4.18 · OpenClaw plugin · MCP coming soon

Decouple your agent's memory from its prompt.

OpenClaw's raw markdown memory is great on Day 1, but breaks down by Day 30. Plumb replaces flat-file injection with semantic retrieval — storing facts as you work and injecting only what's relevant before each response. Already have a MEMORY.md? Plumb seeds from it automatically on first activation.

Read the Docs

or paste this into your OpenClaw chat:

you: Install the Plumb memory plugin…
openclaw — OpenClaw Chat
~0 tokenson cold start
SQLite-veclocal vector search
OpenClaw-nativeplugin in one command
100% privateno cloud required

The Problem

Flat files don't scale. Smart retrieval does.

Every token you spend loading stale context is a token you can't spend solving the actual problem.

MEMORY.md
~12,400 tokens injected
# MEMORY.md — Agent Long-Term Memory
 
## Critical Operational Rules
- Primary channel: Slack. Telegram = fallbackwasted
- Reminders: post to #reminders channelwasted
- CC team@example.com on external emailswasted
- LinkedIn: NEVER send without explicit approvalwasted
 
## Sub-Agent Defaults
- Timeout: 600s (900s for browser tasks)wasted
- Model: claude-haiku-4-5 for sub-agentswasted
- Default: anthropic/claude-haiku-4-5wasted
 
## Key Tool Locations
- Notion task DB: a50b0f76-4900...wasted
- Career facts: career_facts.mdwasted
- Full memory: memory/MEMORY_FULL_BACKUP.mdwasted
... 847 more lines ...
↳ 98% of these tokens are irrelevant to your current task.
plumb.db — retrieved facts
~340 tokens injected
TierSessionAgeFact
HIGHmain-session2 days agoPrimary comms channel is Slack, not Telegram
HIGHmain-session2 days agoSub-agent default timeout is 600 seconds
HIGHmain-session3 days agoCC team@example.com on all outbound emails
MEDsetup5 days agoLinkedIn messages require explicit approval
MEDsetup5 days agoNotion tasks DB id: a50b0f76-4900-431e
LOWsetup6 days agoDefault sub-agent model: claude-haiku-4-5
↳ Only facts relevant to this query were retrieved.
12,400 tokens→ stale, full-file injection340 tokens→ exact-match retrieval via Plumb97.3% reduction

How It Works

Invisible until you need it.

Plumb runs entirely in the background. Install once, forget it exists, then wonder why your agent suddenly remembers everything.

01

Install once

Paste the install prompt into your OpenClaw chat. Your agent handles the rest — download, config, and gateway restart. One conversation, done.

you: Install the Plumb memory plugin. openclaw: Running: openclaw plugins install @getplumb/plumb openclaw config set plugins.slots.memory plumb openclaw gateway restart ✓ Installed. I can feel my brain growing.
02

Your agent writes facts as it works

No prompting required. As your agent learns things worth keeping, it calls plumb_remember() — Plumb handles embedding and indexing in the background. It can also search its own memory mid-conversation with plumb_search().

# Writing a new fact: → plumb_remember("Primary comms channel is Slack", confidence="high") ✓ stored · [HIGH] ready for retrieval # Searching mid-conversation: → plumb_search("sub-agent model") ← Default sub-agent model: claude-haiku-4-5 [HIGH] · 3 days ago
03

Relevant facts injected before each prompt

Before every response, Plumb queries the vector store for facts relevant to the current conversation and injects them as a compact block — never the whole file.

[PLUMB MEMORY — 6 facts, 340 tokens] [HIGH] Primary comms channel is Slack [HIGH] Sub-agent timeout default: 600s [MED] Default model: claude-haiku-4-5 ...
04

Your agent just knows

No new tools. No workflow changes. Your agent simply has context it didn't have before — including things you mentioned in sessions weeks ago.

you: what channel should I post the alert to? openclaw: Slack — specifically #alerts. You set that up a couple weeks ago.

Why Plumb

Built for people who actually use their agents every day.

No magic. No black boxes. Just structured memory that behaves exactly like you'd expect.

Stop paying to read your own chat logs.

Injecting flat files destroys your context window and spikes API costs. Plumb writes high-signal facts as your agent works and injects only what's necessary — semantic search, not grep.

Memory that actually scales past Day 30.

Flat-file memory degrades rapidly as context grows. Plumb ensures an agent with 10,000 memories is just as fast and focused as one with 10. SQLite + sqlite-vec means O(log n) retrieval no matter how long you use it.

Already using MEMORY.md? You're already set.

On first activation, Plumb reads your existing workspace .md files — MEMORY.md, daily logs, notes — and seeds the vector store automatically. Zero migration work. Hundreds of facts from day one.

Stateless agents. Stateful memory.

Your memory lives in a portable SQLite database — not baked into your config files. Back it up, move it between machines, or upgrade to Plumb Cloud for zero-setup cross-device sync.

Plumb Cloud

One brain across every tool. (Coming soon)

Switch from OpenClaw to Claude Code, or move between machines — all sharing the same memory store via MCP. Cross-tool sync is on the roadmap as a Plumb Cloud feature.

Structured like a DB. Readable like a file.

View, edit, and delete any stored memory via the CLI or local UI. Facts are plain-text and human-readable by design — no opaque embeddings, no black boxes. You are always in complete control.

Pricing

Simple. Transparent. No gotchas.

Start free, stay free. Upgrade only if you need sync or want to skip the local setup.

Local / OSS

Tinkerers & privacy hawks

FreeForever

Run Plumb entirely on your machine. Your facts never leave your disk. Full MCP support, zero dependencies.

  • Local SQLite DB (sqlite-vec)
  • Unlimited memory storage
  • Full MCP server support
  • CLI interface (plumb list, plumb wipe, etc.)
  • Auto-seed from MEMORY.md on first activation
  • OpenClaw plugin (memory slot, before_prompt_build)
  • Your memory never leaves your machine
Early Access

Plumb Cloud

Power users & multi-device setups

$9/ month

Everything in the free tier, plus real-time cross-device sync and a hosted web UI to manage your memory store.

Get Early Access
  • Everything in Local / OSS
  • Cross-device real-time sync
  • Hosted vector DB (zero local setup)
  • Web-based memory manager UI
  • Conflict-free merge (CRDT-backed)
  • API access for custom integrations
  • Priority support

OSS core will always be free. No credit card required. Cloud is early access — pricing may change.