Skip to content
MeshTale
Works with Claude, ChatGPT, Cursor — and whatever you use next

One memory for every agent. Never the wrong client.

Your people and your agents produce context all day. MeshTale organizes it, then decides who gets which part back.

Producing raw context

DanaClient calls
Claude CodeSession notes

One governed memory

MeshTaleapproval · audience · retention
Northwindworkspace
Client commsContracts
Ardent Coworkspace
BriefsDelivery
Kestrelworkspace
SupportSOWs

Pulling the right context

Claude CodeDrafting the renewalNorthwind only
SamAnswering the clientNorthwind only

Free tier, no card required. Source-available under FSL-1.1-MIT.

What that looks like

Ask in the tool you’re already in

No new window, no copy-paste, no re-explaining the account. Your agent asks, MeshTale returns what that client’s workspace holds — and shows its working.

Scoped
Only the workspace that agent is bound to.
Cited
Every answer names the sources it drew on.
Logged
The recall is recorded, whoever made it.
Claude Codenorthwind

ask your own memory…

Numbers we can actually show our working for

No hours saved, no percentage lifts, no months to payback. We have no customers yet, so we have no business claiming any of that. These are properties of the product, measured or counted.

0

tools to register

Most setups spend tens of thousands of tokens describing their tools before the model does any work.

0

tokens sent per call

Measured, not estimated — both tool definitions serialised and counted. Reproducible.

~0

permission checks per answer

Stacked authorization filters, applied to every query on every route.

0

sources you can connect

Real implementations against real vendor APIs, not a roadmap slide.

0%

retrieval recall

On a public memory benchmark, 311 questions. Measures whether the right material was found — not end-to-end answer accuracy.

0

model providers, your keys

Bring your own OpenAI, Anthropic or Google key and run your memory against whichever you prefer.

How it stacks up

Your team and your agents, drawing on the same place

Your people produce the context — calls, briefs, specs, contracts. Your agents consume it. Both go through the same workspace, sorted into collections, under the same permission checks. One workspace per client, and nothing crosses between them.

Your organization
Northwindworkspace
Client comms
Agreed pricing floor
Kickoff decisions
Contracts
MSA v3
Kestrelworkspace
Support
Escalation policy
SOWs
Phase 2 scope

Same organization · no path between the two workspaces

  1. 01

    Organization

    Your company. Holds every workspace and the people in them.

  2. 02

    Workspace

    Usually one per client or team. This is where the boundary lives.

  3. 03

    Collection

    A folder inside a workspace, with its own sensitivity default.

  4. 04

    Memory

    A single piece of recallable context, carrying its own audience tags.

Why it’s different

01 · Separate by default

One client’s context never reaches another’s prompt

Tenancy runs through the data model rather than sitting on top of it. Audience tags travel with the content, so the boundary is a property of the memory itself — not a filter someone remembered to apply.

  • Row-level isolation in the database
  • Per-key scoping, bound to a workspace
  • Audience tags on every piece of content
Memory · NorthwindApproved

Renewal pricing agreed at $4,200/mo, effective from the March invoice.

audience: account-teamsensitivity: confidential#northwind-billing
Claude Code · Northwindreturned
Cursor · Ardent Conot visible

02 · Answerable

When someone asks what the AI saw, you have a record

Every recall is logged — which agent, which workspace, how much came back, and when. Not a summary written afterwards, but the thing itself.

  • Query log, security events, citation graph
  • Answers cite back to the source
  • We log what was returned, not what was used
Query log · liverecording
  • claude-coderecalled 3 memories from Northwind14:02
  • cursorrecalled 1 memory from Kestrel14:02
  • api · reportingrecalled 8 memories from Northwind13:58
  • claude-coderecalled 2 memories from Ardent Co13:51
  • chat · danarecalled 5 memories from Kestrel13:44

every recall · who, what, when, which workspace

03 · Cheap to live with

Two tools, one sentence each

A typical setup spends tens of thousands of tokens describing its tools before the model does any work. Ours registers two, and every answer comes back with its sources attached.

  • Two registered tools, not dozens
  • Works with any client that speaks the protocol
  • No software library to adopt, nothing to unpick

What did we agree with Northwind about the renewal?

Renewal was agreed at $4,200 a month, starting from the March invoice, with the extra seats folded in rather than billed separately.

Drawn from
  • #northwind-billingSlack · 12 Jun
  • Northwind MSA v3.docxDrive · 14 Jun

Northwind workspace only · nothing from another client

The product

Governance you can point at

Filter memories by who may see them, how sensitive they are, and where they came from. Retention and policies sit in the same navigation, not in a slide about next quarter.

Audience Access
who may see what
Sensitivity Labels
how guarded it is
Retention
how long it stays
Policies
what the rules are

Agreed pricing floor for renewal

slack #northwind-billing · 12 Jun

Audience
Northwind workspace
Sensitivity
Restricted
Retention
24 months
Status
Approved

Who reaches it

  • Claude Codeworking in Northwindallowed
  • Cursorworking in Kestrelout of scope
  • Support assistantcustomer-facingout of scope

The refusals are not a filter applied to the answer. Those two never had this memory in scope to begin with, and the attempt is recorded either way.

How it works

Three steps, and your agents stop asking you to re-explain

  1. 01

    Connect your sources

    Point MeshTale at the places your work already lives. It parses, chunks and indexes.

  2. 02

    Set the boundaries

    Decide who and what can see each piece. Audience tags travel with the content.

  3. 03

    Recall from any agent

    Connect Cursor, Claude Code, or any MCP client. Ask; get cited answers scoped to what that agent is allowed to see.

✓ Connected. 2 tools registered.
  recall · recall_context_bundle
↳ 3 memories · Northwind workspace only
  slack #northwind-billing · 12 Jun
  drive / Northwind MSA v3.docx
✓ Cited. Logged. Nothing outside this client.

Underneath

The parts nobody puts on a homepage

Storing text is the easy half. These are the decisions that determine whether what comes back is any good — and whether you can live with it once more than one person is involved.

Take some of it, not all of it

Choose the folders and channels worth bringing in rather than hoovering up an entire account. Fewer irrelevant results, and a smaller surface to govern.

  • #northwind-billing
  • /Northwind/Contracts
  • #random
  • /Personal

Enforced today on Agency and Scale

Not every memory carries equal weight

Content arrives tagged by how much it should be trusted, and that tag feeds the ranking — so an approved contract outranks a passing remark in a thread. Your organization sets the rules that assign those tags.

  • 1Signed MSA v3Approved
  • 2Pricing decision, #billingVerified
  • 3"I think we said 12%"Unverified

Ranked by trust, not just by similarity

Trust tags feed the ranking score

Your own brain, and the company one

Personal connections stay yours — your mail, your notes, visible only to you. Shared connections belong to the workspace. Both live in the same product without mixing.

Personal

Your mail
Your notes

visible to you

Shared

Client drive
Team channels

visible to the workspace

Two separate connection lanes

Two ways of searching, merged

Exact-match and meaning-based search run separately, then get combined by rank rather than by raw score — which stops one drowning out the other when they disagree.

Exact

  • 1MSA v3
  • 2Billing thread
  • 3Renewal notes

Meaning

  • 1Renewal notes
  • 2MSA v3
  • 3Kickoff call

Merged

MSA v3 · Renewal notes · Billing thread

Reciprocal rank fusion

Newer versions of a source supersede older ones, and recency feeds the ranking. A full decision history — what changed, when, and why — is not built yet, so we are not showing you one.

Ways in

However you work, there’s a door

The same governed memory behind every route — with the same permission checks on each one.

Model Context Protocol

The open standard your AI clients already speak. Two tools to register, one sentence each — so connecting costs almost nothing before you have asked anything.

  • Two tools to register, so the setup cost is one command
  • Answers come back scoped to the workspace that agent is bound to
  • Every recall is cited and logged, whichever client asked
claude mcp add meshtale
  • Claude Code
  • Cursor
  • Codex
  • Gemini
  • Grok
  • Manus
  • Hermes
  • OpenClaw
  • Windsurf
  • Zed

…and anything else that speaks the protocol

The honest bit

Couldn’t I just build this myself?

Yes. And it’s easier than people admit — free, well-documented projects stand this up in well under an hour, with a community shipping importers for every tool you use.

If you’re one person organising your own thinking, that is genuinely the right answer, and we’d rather you spent nothing. We are not going to tell you self-hosting is hard, because it isn’t.

Then you take on a second client.

A personal memory store has one owner by design. The moment you’re holding two people’s work, four questions arrive — and none of them are weekend problems.

  1. 01Which client’s context just went into that prompt?
  2. 02Can this agent see the engagement it isn’t working on?
  3. 03When they ask what the AI was allowed to read, what do you show them?
  4. 04When someone leaves, what exactly do you revoke?

Isolation, permissions and an audit trail aren’t features you bolt on later. They’re the reason this exists.

Plugs into what you already use

All 26 of them, on every plan. What changes higher up is how precisely you can choose which folders and channels come in.

MeshTale connects to 26 sources: Slack, Microsoft Teams, Discord, Gmail, Outlook, Notion, Confluence, Google Drive, Dropbox, SharePoint / OneDrive, Obsidian, Microsoft OneNote, Amazon S3, Jira, Linear, Asana, ClickUp, Trello, Monday.com, GitHub, Salesforce, HubSpot, Zendesk, Granola, Fathom, Zoom.

Stop re-explaining your work to every agent

Without leaking one client’s context into another’s prompt.

Free tier, no card required. Source-available under FSL-1.1-MIT.