Memoryextracted from Linear ENG-231 · 6 hours ago
“The billing vendor caps us at 60 requests a minute.”
For teams running AI agents
One shared brain for every agent your team runs. Harbor reads your Slack, repos and docs, hands each agent the few facts that matter, and takes back what the session learned.
Everything your team already produces, read continuously and turned into one graph. Nobody opened a document to write any of this.
Jira, GitLab, Confluence, Gmail, Outlook, 14 repos, 38 uploads · all live
0 facts · 0 connections
updated 40 seconds ago
Nothing enters the graph until a person says yes.
“The billing vendor caps us at 60 requests a minute.”
“Prod deploys go through the staging gate. No exceptions.”
“…so the gate is mandatory from today. No direct-to-prod, ever, not even for a one-line fix.”Dana Whitfield · #eng-releases · 12 August
“Every migration ships with a down script.”
11 more waiting ·j/k move · a approve · d dismiss · u undo
How can I help?
Ask what your team knows: conventions, decisions, and memory.
Why do we deploy through staging first?
Today 09:24
Why do we deploy through staging first?
deploy.yml refuses any ref that has not passed staging.3Ask, search, or make anything…
Nobody pastes context into a prompt. The CLI, the desktop app and the MCP server put the graph inside the run, and carry the run back.
Harbor CLI
harbor claude · codex
Desktop app
menu bar · always on
Harbor MCP
harbor-mcp · stdio
And the road runs both ways.What context got used, what got ignored, what a person overruled: the same three carry it home, and the graph changes. Nobody files a report.
1,492 runs read back todayOne layer, two doors. Your agents read it as files. Your team reads it as an answer. Nobody wrote it down.
The loop
Nobody writes a document for this. Harbor reads the threads, pull requests and calls where the decisions already happen.
Your company
Knows everything.
Your agents
Start already knowing.
Your agent's memory is yours alone. This one belongs to the team, and it fills up from everybody's sessions, not just yours.
The run
No search box, no paste, no command. What your team knows is already in the session. So is every skill they wrote.
Engineering
claude "add the refund endpoint"
Without Harbor
→assumed a 30-day refund window
→called the billing API inside the request handler
Wrong policy, shipped. Caught in review two days later.
claude "add the refund endpoint"
With Harbor · 2 facts, picked for this run
→60 days on annual plans, 30 on monthly#support-leads
→queued the billing call, never inlinePR #3187
Merged on the first pass. No review comments.
Sales
claude "draft the renewal quote for Northwind"
Without Harbor
→offered 28% off to get it signed this week
→promised a February start on an annual term
Margin blown. Re-cut by finance before it went out.
claude "draft the renewal quote for Northwind"
With Harbor · 2 facts, picked for this run
→held the discount at 20% without TomasPricing call, 4 Aug
→started the term on signature, not on kickoffRevenue policy
Sent the same day. Approved exactly as written.
Support
claude "write the macro for enterprise refunds"
Without Harbor
→told the customer 30 days, flat
→routed the ticket into the general queue
Contradicted the policy doc. Escalated within a day.
claude "write the macro for enterprise refunds"
With Harbor · 2 facts, picked for this run
→split the window by plan, 60 and 30#support-leads
→sent enterprise tickets to Dana firstSupport routing
Shipped as written. No escalation, no correction.
Same model. Same prompt. The only difference is what it already knew: two facts out of 696, picked for this run.
Written once, never pasted anywhere. A skill or a document is attached to a space, and the space decides where it lands.
Written once
Release checklistskill
Incident write-upskill
Refund policy v4document
7 skills · 12 documents
Attached to a space
Engineering14 repos · 24 members
Support2 repos · 11 members
Everyoneevery member
3 spaces in this workspace
In every environment it owns
~/.<agent>/skills/every agent
.claude/docs/acme/api +2
CLAUDE.md · AGENTS.md1 line
written 2 minutes ago
Nobody chooses a space. acme/api is bound to Engineering, so a run that starts in that repo is already in it. A skill needs no repository at all, and a document of any size costs the session one pointer line until an agent opens it.
You already have a file that does part of this. Two things stay broken however long you make it.
One audience, and it is everyone
Whoever opens it reads all of it. Support gets the deploy rules, engineering gets the refund policy, the contractor gets both. Keep one file per team instead and you are maintaining four that have already started to disagree.
No way to say a line went stale
deploy/staging.yml was deleted in June. The rule that names it is still in the file, and still being read. Nothing about a file changes when the thing it describes stops existing.
Neither is a length problem, so neither is fixed by a shorter file. Or a longer one.
Four windows onto the same layer, and you already have one of them open.
npm install -g harborloop
harbor init
Once, on the machine. Claude Code, Codex and Cursor all read it.
harbor_search_knowledge
Six read tools, and every answer comes back with a [#1] on it.
/harbor [in Engineering]
Answered in the channel, to you alone, with the [#1] under it.
Review · 3 waiting
Ten minutes a week, and the count of what your agents used.
Control
Every proposal arrives with the thread it came from, so you can read what was actually said before you say yes.
Dana Whitfield Tue 10:04
Enterprise is asking about the refund window again. What do we actually tell them?
Marcus Webb Tue 10:07
I thought it was 30 days flat? That's what I've been sending.
Elena Novak Tue 10:09
Refunds run 60 days on annual plans, 30 on monthly. It has been that way since March.
Priya Lal Tue 10:11
Confirmed. That's what the billing policy doc says too.
Marcus Webb Tue 10:12
Then I owe three people a correction. Sending them now.
Dana Whitfield Tue 10:14
Can we put this somewhere the new folks will actually find it?
35 more messages
Refunds run 60 days on annual plans, 30 on monthly.
Elena Novak Tue 10:09
Support
ApproveNot a rule
Staging deploys run from deploy/staging.yml, never from the release workflow.
deploy/staging.yml was deleted. Still being served.
Engineering
The billing vendor caps us at 60 requests a minute.
Engineering
Nothing is read until you pick it.
Approve everything
Nothing reaches an agent until a person says yes. Where every workspace starts.
Hybrid
Obvious things go straight in. Anything load-bearing stops and waits for you.
Autopilot
Straight to the graph. For a space that has watched the queue long enough to trust it.
A file on disk never arrived with the thread that produced it, and never told you when it stopped being true. Neither does a search box pointed at the same files.
Ask
Anyone can ask, in Slack or in the app. Every claim carries the sentence it came from, opened in place.
How can I help?
Ask what your team knows: conventions, decisions, and memory.
Enterprise customer, annual plan, day 45. Do we refund?
Today 09:24
Enterprise customer, annual plan, day 45. Do we refund?
Searching team knowledge
Ask, search, or make anything…
What it proves
Counting how often a rule reached an agent is free with your model vendor now. Whether the answer used it is a different question, and the field’s own survey, over 400 works, still files that one under open problems.
The numbers are theirs, not ours. Over 138 tasks and four agents, a hand-written context file moved task success about four points, a generated one moved it down, and either way the run cost about 20% more inference (ETH Zürich and LogicStar, arXiv:2602.11988). Which is the objection to us too, so Harbor reads its lines out of work people already did and a person approves every one before an agent sees it. The open-problems survey is Always-On Agents, arXiv:2606.30306.
Your 238 rules. Harbor served 91 of them this week.
34
Came back cited
Named as what changed the answer.
57
Still on trial
Sent this week, not yet named.
Agents only count if they tell us. 310 of 402 sessions reported back, so these are floors, never ceilings.
| Most served | What it did |
|---|---|
| Queue the billing call, never inline | Changed 41 answers |
| Refunds: 60 days annual, 30 monthly | Changed 31 answers |
| Prefer named exports over default | 84 handovers, never cited |
| Run the full suite before pushing | 76 handovers, never cited |
147 rules never matched a task in 402 sessions.
See every rule, and delete the ones nobody uses→
Start here
Nothing to write, and nothing to migrate. Harbor reads the repos, threads and calls your team already produces. Two weeks of real sessions later, every rule carries a served count, the ones an answer named carry a cited count, and the rest are the list you can delete.
You will talk to the people building it, not to a sales team.