Two small tools, not a toolbox
MeshTale adds two tools, described in under 1,000 tokens. Many setups spend tens of thousands before the model does any work.
Your people and your agents produce context all day. MeshTale gives each of them the right context for the client they’re on, and never the wrong one’s.
Free tier, no card required.
Each client gets its own workspace. An agent bound to one workspace never has another client's memories in scope, and every attempt to reach them is logged.
The refusal is not a filter on the answer. The out-of-scope memory was never in that agent's reach, and the attempt is recorded either way.
Every memory sits in a collection, inside a workspace, inside your organization. A recall passes each layer on the way in, and the workspace wall is where the wrong client gets stopped.
A stack of four layers. The organization is the base. On it sit three walled workspaces: Northwind, Ardent Co and Kestrel. Inside Northwind are three collections: Client comms, Contracts and Delivery. Memories sit on top, each showing who created it: people such as Dana, Marco and Priya, or AI tools such as Claude Code, Codex and Cursor.
Northwind / Contracts
by Marco · Granola call
Cited in Claude Code's answer
Cursor is bound to Kestrel. Northwind is out of its scope.
Plug it into the assistant you already use, call it from your own code, or sync coding-agent sessions from the command line. Same scoping and the same log, whichever door you use.
No new window and no re-explaining the account. The agent asks, MeshTale answers from one workspace, cites the sources and logs the recall.
Tela, the chat inside MeshTale, searches the sources connected to a workspace and cites the memory behind each sentence. Point at a number to see it.
Northwind chat
Ask Tela
What do you want to know about Northwind?
Tela answers from your connected sources with citations.
Press Enter to send
Reading workspace memory…
Northwind gets a two-week guided onboarding, starting Monday 2 November. Priya runs the kickoff in week one and opens a shared Slack channel for their team.
Moving data from their old CRM covers contacts and deals only, not email history. A named support contact for the first 30 days is on the welcome checklist.
Dana: We run a two-week guided onboarding, starting Monday the 2nd. Priya owns week one: the kickoff call, then a shared channel so their team can ask us anything.
Captured 8 days ago
Priya: Kickoff is booked for week one. I'll open a shared channel with their ops team the same day.
Codex: Import scope follows the SOW: contacts and deals only. Email history is excluded, so the script skips the activities table.
Claude Code: Added step 6 to the welcome checklist: a named support contact for the first 30 days after go-live.
Your assistant calls MeshTale only when it needs context, and gets back just the cited pieces that answer the question. Your session stays fast and keeps its room for the actual work.
MeshTale adds two tools, described in under 1,000 tokens. Many setups spend tens of thousands before the model does any work.
Each recall comes back within a size budget: the most relevant cited pieces first, never whole documents pasted into your chat.
Ask the same thing again and the answer comes straight from cache, still scoped to the same workspace and still logged.
Who may see a memory, how guarded it is, how long it stays and whether a person has approved it all travel with the memory itself.
Personal data such as emails, phone numbers and ID numbers is detected on the way in and can be masked before it ever becomes a memory. Each workspace sets its own data policy and restrictions.
New memories can wait for a person to say yes before any agent can see them. Trust rules auto-approve the sources you already trust, so the queue only holds what needs a look.
Audience tags ride on the content, so the boundary belongs to the memory itself.
Who asked, where, and what came back, and when. The record itself, not a summary.
Every collection has a default, and any memory can be marked more guarded than its neighbours.
Decide how long each kind of context stays, in the same place you set who sees it.
Connect your own inbox for yourself, or a team channel for everyone in the workspace.
Keyword hits and similar-meaning hits, merged into one ranked list.
MeshTale notices when a new memory contradicts one you already hold, and asks the right person which one is accurate. Answers follow their decision.
"Onboarding is two weeks, with a named contact."
"We agreed three weeks of onboarding."
Onboarding length: 2 weeks vs 3 weeks
Agent answers about Northwind will follow whatever Priya decides.
Onboarding is two weeks, with a named contact.1
Onboarding is three weeks, as agreed at kickoff.1
Three weeks. The SOW says two, and Priya notes the kickoff added a week.1,2
When a new memory contradicts one already in the workspace, both are marked and nothing is silently overwritten.
The alert goes to the workspace's reviewer. They keep one version, or keep both with a note that explains the difference.
The decision lands in the audit log, and agents cite only the version the person confirmed from then on.
Connect the places work already happens. MeshTale reads, sorts into the right workspace and hands the right part back out.
All 26 sources on every plan. Browser extension in development.
A command-line sync tool reads Claude Code, Codex, Gemini CLI, Grok and Muse session histories into the right workspace. Decisions, dead ends and fixes stop evaporating when a session closes.
Most setups spend thousands of tokens describing their tools before the model does any work. MeshTale registers two.
Stacked authorization filters, applied to every query on every route.
Real implementations against real vendor APIs, plus webhooks in and out.

One workspace per client, so the brief for one never leaks into another.

Engagement history every consultant and agent can draw on, kept apart.

Several companies in one head. Now in separate workspaces too.

Coding agents that remember each client's codebase decisions.

Every account's calls and threads, ready before the next one.

One memory across functions, with audience rules per team.
Local memory is great for one person on one computer. A homemade setup can do anything, if someone builds it. Here is where each one stops.
| What you need | Local memoryNotes files or a memory tool on your computer | Do it yourselfA vector database and a prompt | MeshTale |
|---|---|---|---|
| Works without signing up for anything | Yes | Yes | Free account |
| Runs entirely on your own computer | Yes | If you host it | Hosted, export any time |
| Store and search context | Your own | Yes | Built in |
| Shared by the whole team, on any device | One person, one machine | Build it yourself | Built in |
| Works in every assistant, not just one | Usually one tool | Build it yourself | Built in |
| One workspace per client, enforced in the data model | Separate folders, by habit | Build it yourself | Built in |
| Every recall logged with agent, workspace and result | No | Build it yourself | Built in |
| Answers that cite their sources | Sometimes | Build it yourself | Built in |
| Audience, sensitivity and retention per memory | No | Build it yourself | Built in |
| Review queue before agents can recall | No | Build it yourself | Built in |
| Personal data detected and masked before storing | No | Build it yourself | Built in |
| 26 connectors and coding-agent session sync | Copy things in by hand | Build it yourself | Built in |
Try it with your own agents. No card required.
For one person working across several clients.
A team, many client workspaces, review and audit.
More workspaces, more sources, more agents.
Custom terms and a direct line to the team.
Each client usually gets its own workspace, and every key and connection is bound to one. Separation runs through the data model, so a memory outside an agent's workspace is never in scope for it. Attempts to reach it are logged.
Anything that speaks MCP, including Claude Code, Cursor, Codex, Gemini, Grok, Windsurf and Zed. You can also call the REST API with a workspace-scoped key, or use the Obsidian plugin.
It is detected as content comes in, and each workspace can choose to mask it before it is stored. Data policies and restrictions are set per workspace.
Yes. New memories can sit in a review queue until someone approves them. Until then, no agent can recall them.
It is source-available under FSL-1.1-MIT, and there is a free tier with no card required.
Connect a source, connect an agent, and ask your first question.
Free tier, no card required.