Skip to content
How it learnsFrom coding sessions, reviews, Slack and docs.Only what fitsEach session gets the few rules its task needs.Pull request reviewEvery PR checked against your team’s rules.Keep it trueA gate, conflicts, disputes and stale rules, handled.HandoffsCarry the work to another session or a teammate.SecurityWhat it sends, what it changes and how to leave.
Engineering teamsOne agent or several: what one person learns, every session uses.PR review onlyInstall the GitHub App. Nothing on anyone’s machine.Teams that mix agentsClaude Code, Cursor, Codex and Copilot, the same rules.On your ownYour corrections follow you across repos, machines and agents.Book a call30 minutes with Roei, if you want a hand. Never required.
DocsPricingSign in
Connect your agents →
For teamsOn your ownPR review onlyHow it learnsOnly what fitsPull request reviewKeep it trueSecurityDocsPricingSign inConnect your agents →
Home/Blog

Blog

AGENTS.md, Cursor rules and Copilot instructions

  • AGENTS.md vs CLAUDE.md
  • Sync rules across repos
  • Why Codex ignores it
  • Keep rule files in sync
  • 3,905 files checked
  • Why Copilot ignores it
  • Cursor Team Rules
  • Cursor rules
  • copilot-instructions.md
  • Claude Code and AGENTS.md
  • AGENTS.md research

CLAUDE.md and context engineering

  • Why Claude ignores it
  • Test a CLAUDE.md rule
  • One CLAUDE.md for all projects
  • Context engineering, Claude 5
  • CLAUDE.md length
  • Which rules got used
  • Context rot, stale rules
  • Growing CLAUDE.md
  • Claude Code best practices
  • Karpathy CLAUDE.md
  • Committing CLAUDE.md

Built-in agent memory

  • Agent memory compared
  • Claude Code auto memory
  • GitHub Copilot Memory
  • Make Claude Code remember
  • Claude Tag vs Claude Code
  • Codex memories
  • Cursor memories
  • Memory or documentation?

Team knowledge and shared conventions

  • Claude Code team memory
  • Who owns CLAUDE.md
  • Claude Code second brain
  • Shared memory across agents
  • Stop repeated mistakes
  • Hand off between agents
  • Company brain
  • Slack decisions
  • Memory poisoning
  • Karpathy's LLM wiki
  • ADRs for agents

Skills, hooks and permissions

  • Rules vs skills
  • Claude Code hooks
  • Codex skills
  • Auto mode
  • CLAUDE.md vs skills vs hooks
  • Claude Code skills

AI code review and PR guardrails

  • PR review against AGENTS.md
  • Review comments to rules

Memory tools and alternatives

  • Best company brain tools
  • Unblocked alternatives
  • claude-mem alternatives
  • Supermemory alternatives
  • Mem0 alternatives
  • claude-mem
  • Best MCP memory servers
  • Supermemory Company Brain alternatives

Blog

The best company brain tools for coding agents (2026)

Published October 1, 2026 · Updated November 17, 2026

We make Harbor, and it is first on this list. Every other entry was checked against its own public pages on 1 October 2026, with the review steps checked again on 5 October. MemoryLake, Sentra and Coworker were added on 15 November, and ByteRover was read again on 17 November after it changed what it sells. Each one says where it does something better than we do.

The short answer: if your team's decisions live in Slack threads and pull request reviews and you want them in every coding session, with a count per rule of how often it was served and how often an answer cited it, that is Harbor. If you want answers on demand from every system you run, with enterprise controls today, look at Unblocked. If you want one governed playbook written into every agent's own rule files, look at Packmind. For open-source memory you host, Cognee. For search over the documents you upload, with the passage and its page, ByteRover. For the fastest automatic memory in Claude Code, Mem0's plugin. For repo memory that branches and merges with your pull requests, MemoryLake. For one memory of the whole company, sales and finance included, that agents can also read, Sentra or Coworker.

What counts as a company brain for coding agents?

A company brain holds what a company decided and hands the relevant part to whoever is doing the work. For an engineering team in 2026, the busiest reader is the coding agent, so this list only includes tools that Claude Code, Cursor or Codex can actually read. We judged each one on eight questions:

  • Where the rules come from. Written by hand, saved by agents, or learned from where the team decides.
  • Who decides what gets in. Automatic, a review step, or a gate you set.
  • How it reaches the agent. Files in the repo, context added at the start of a session, or a tool the agent has to think to call.
  • Only what applies, or everything. Whether the agent gets the rules for its task or the whole file.
  • What happens when a rule goes stale. Whether anything notices that the code moved on.
  • Whether you can see what got used. Not what was stored or sent, but what an answer used.
  • Whether it can stop an action. A rule the agent reads is advice; a block is a guarantee.
  • Agents, licence, price and availability. Including whether you can start today.

Start with what you have: CLAUDE.md, AGENTS.md and Cursor rules

Before any tool, there are the files every agent already reads. A CLAUDE.md holds instructions Claude Code loads at the start of every session. AGENTS.md is the shared format that Codex, Cursor, Copilot, Antigravity CLI and about twenty other tools read, and Claude Code reads it too when a project has no CLAUDE.md. Cursor adds .cursor/rules, where each rule can apply always, by file pattern or on request.

They are free, they live in git, and for a small team they are the right place to start. Their limits show up as the team grows. Someone has to write every line, every agent gets the whole file, nothing notices when a line stops being true, and Anthropic's own docs describe them as context rather than enforced configuration. We have written up why a growing CLAUDE.md stops working and how AGENTS.md and CLAUDE.md differ. Every tool below is an answer to one or more of those limits.

1. Harbor

Harbor learns the rules your team decides in Slack, pull request reviews and docs, and gives each Claude Code, Cursor, Codex or Copilot session the ones that apply to its task before the agent writes a line. What a session finds out goes back in, and you set the gate: Manual holds everything for a person, Hybrid (the default) puts routine rules live on trial and holds the rest, and in every mode a secret, an install step or a clash Harbor cannot settle waits for a person.

  • Best for: engineering teams whose decisions live in threads and reviews, and who want every agent to start with them.
  • Rules come from: Slack, GitHub, GitLab, Notion, Confluence, Jira, Linear, Google Drive and Granola, plus what agent sessions record.
  • Reaches the agent: hooks add the rules that apply at the start of each session. The same rules are on an MCP server, the CLI and /harbor in Slack.
  • Where it is ahead: every rule shows how often it was served and how often an answer cited it, so you can delete the ones nobody uses (served vs cited). Block rules refuse a destructive command before it runs. When a push deletes the file a rule is about, the rule is retired, or flagged if other repos still use it.
  • Watch out for: It is new: public sign-up opened on 31 October 2026, self-serve, with no card and no call. It wires four agents, not ten. The service is hosted only, and there is no SSO or enterprise plan yet.
  • Price: Free costs nothing for up to 2 people, on one source. Pro is $24 per person a month billed yearly ($30 monthly), and Premium $36 ($45). The CLI, harborloop, is MIT licensed. Every limit is on the Harbor pricing page.

2. Unblocked

Unblocked is a context engine for engineering teams. It reads about 45 sources, from code and pull requests to Slack, Teams, docs, issue trackers, CI and runtime tools, reconciles them and answers people and agents with citations. It now also learns conventions it sees repeated in pull request review feedback.

  • Best for: larger engineering organisations that want answers on demand across many systems, with SSO, SCIM and SOC 2 Type II now.
  • Who decides: nobody approves rules. Anyone in the organisation can correct or remove a shared memory.
  • Reaches the agent: the agent calls its MCP tools or CLI when it decides to, or when you ask it to. Its docs suggest a line in your CLAUDE.md telling agents to use it.
  • Where it is ahead: the widest source coverage here, conflict handling across sources, and the most mature enterprise controls. It also sells code review and a CI failure agent.
  • Watch out for: nothing is in the session until the agent asks, analytics count tool calls rather than which rules were used, and there is no free plan.
  • Price: $29 per user a month billed yearly, $35 monthly, after a 21 day trial.

The two overlap more than any other pair on this list. Read Harbor vs Unblocked for the rows side by side, or Unblocked alternatives if you are leaving it.

3. Packmind

Packmind keeps one engineering playbook of standards, commands and skills, and writes it into each agent's own files across your repos: .claude/rules, .cursor/rules, Copilot instructions, AGENTS.md and more.

  • Best for: organisations with many repos and several agent tools that want one versioned source of truth for their rule files.
  • Who decides: edits go live directly, or as Change Proposals that the team reviews before they apply. That review workflow is the closest thing on this list to a gate, and it is on the Enterprise plan.
  • Reaches the agent: as files committed to the repo, loaded whole or by file pattern, the way the native files are.
  • Where it is ahead: renders to ten agent formats, is open source under Apache 2.0, self-hosts, and has an Enterprise linter that catches rule violations before a commit.
  • Watch out for: rules change only when someone proposes a change. Mining Slack and pull requests is done with example skills you run yourself, not inside the app.
  • Price: the open-source edition is free. Enterprise, which carries the review workflow and the linter, is priced on request.

4. Cognee

Cognee is an open-source memory engine that builds a knowledge graph from documents, code and agent sessions. Cognee Cloud adds a hosted Company Brain.

  • Best for: teams that want graph memory they can host themselves, even air-gapped, and builders putting memory into their own product.
  • Rules come from: mostly what agent sessions record. In Cloud, Slack, GitHub and Google Drive are live, and its docs list more sources as coming.
  • Reaches the agent: its Claude Code and Codex plugins add recalled context to every prompt through hooks. Cursor and others use its MCP server.
  • Where it is ahead: Apache 2.0 with a real self-host story, the deepest Claude Code integration of the memory vendors, and a Memory Coverage view that scores how well memory answered.
  • Watch out for: extraction is automatic with no review step, and usage is priced in tokens, so cost grows with how much it processes.
  • Price: free up to 1M tokens, then $1 per 1M tokens and $5 for each extra workspace. Enterprise runs in your cloud.

More in Harbor vs Cognee.

5. ByteRover

ByteRover has changed since this list was first written. It was a memory CLI for coding agents, and as read on 17 November 2026 it is document search for agents: you upload files to a drive, and your agents get back the passage and its page. Its own FAQ says it is search, not a memory. The memory CLI's repository is archived, and its last release was on 27 May 2026.

  • Best for: teams whose agents need the exact passage and page from contracts, policies, specs and other documents.
  • Rules come from: the files your team uploads. It does not learn rules from sessions, Slack or pull requests, and a connected agent saves a file only when you ask it to.
  • Who decides: no review step is documented. Roles decide who can add files to a drive.
  • Reaches the agent: when the agent calls its MCP server. A search takes about three seconds and returns up to 10 passages, each with its file and page.
  • Where it is ahead: it reads PDF, Word, PowerPoint and Excel files, any agent that supports MCP can connect, and it is SOC 2 Type II audited.
  • Watch out for: it is not a memory, so nothing an agent learns in a session is kept unless you upload it as a file.
  • Price: $0.002 a search at medium effort and $0.30 a GB a month past the first GB, with no seats. A first team starts with a $5.00 balance and no card.

More in Harbor vs ByteRover.

6. Mem0's coding plugin

Mem0 is a memory layer for the agents and apps you build. Its plugin for Claude Code, Codex, Cursor and others captures what happens in a session and sorts it into personal memory and a project pool that everyone on the repo reads and writes.

  • Best for: the fastest automatic memory across sessions, and teams already building their product's memory on Mem0.
  • Who decides: extraction is automatic, with no review step documented.
  • Reaches the agent: in Claude Code and Codex, the plugin adds up to five memories when you send a first prompt of 20 characters or more, and the agent can search for more. In Cursor it adds none on its own.
  • Where it is ahead: one install to start, newer facts supersede older ones on every hosted plan, and a large ecosystem of 31 integrations.
  • Watch out for: a wrong project memory reaches the whole repo, it learns nothing from Slack or reviews, and the plugin needs the hosted platform even though the library is open source.
  • Price: free with usage billing past its limits, or a custom plan.

More in Harbor vs Mem0.

7. MemoryLake

MemoryLake sells codebase memory: one versioned memory per repository that Cursor, Windsurf, Claude Code and custom agents all read. It versions memory like code, so what an agent learns on a feature branch stays there and merges with the pull request.

  • Best for: a team that wants repo memory to follow its branches, and to look back at what an agent knew at a given moment.
  • Rules come from: what AI tools write as decisions get made, plus what you import by command: a docs folder, pull request reviews, single facts.
  • Who decides: its codebase page describes no review step.
  • Reaches the agent: over MCP or post-commit hooks, and an HTTP API for custom agents.
  • Where it is ahead: memory branches, diffs between any two points, and queries against memory as it was at a past moment. Deployment in your own VPC is offered, with SOC 2 and ISO 27001.
  • Watch out for: it is memory of a repository, not of where the team decides. Nothing on that page reads Slack threads or shows which memories an answer used.
  • Price: free up to 300,000 tokens a month, then Pro at $19 a month and Premium at $199 a month, priced in tokens.

8. Sentra

Sentra calls itself managed organization memory for teams and AI agents. It reads more than 200 of the systems a company runs on, from email, chat and meetings to the CRM, ERP and HR systems, keeps track of what is true now and why it changed, and flags problems to the right person.

  • Best for: a company that wants one memory across every department, with people and agents reading the same thing.
  • Who decides: nothing Sentra drafts or sends reaches anyone until a named person says yes. That gate is on what Sentra does, not on what enters its memory.
  • Reaches the agent: its MCP server names Claude Code, Codex, Cursor, VS Code, Windsurf and Zed. The agent calls it for meetings, memory and entities. Its docs do not describe adding context at the start of a session.
  • Where it is ahead: the widest company coverage on this list, SOC 2 Type II and ISO 27001, and deployment in its cloud, your VPC, or air-gapped on premises.
  • Watch out for: nothing on its pages is specific to code: no rules from pull request reviews, no per-rule count of what agents used.
  • Price: not published. You book a demo.

More in Harbor vs Sentra.

9. Coworker

Coworker is a context layer for enterprise AI. It connects more than 50 work tools into a company graph it calls OM2, and sells its own apps on top, including Code, a coding agent for the terminal.

  • Best for: an enterprise that wants one context layer for every AI tool, and may run Coworker's own coding agent.
  • Rules come from: 50+ connectors, plus custom connectors for internal systems.
  • Reaches the agent: any MCP client, and its pricing page names Claude Code and Cursor. It does not describe how context reaches a session.
  • Where it is ahead: one layer for every AI tool in the company, with SOC 2 Type II, GDPR and CASA Tier 2.
  • Watch out for: one Enterprise plan, sold after a demo. Nothing on its pricing page covers rules from reviews, a gate on what agents are told, or what an answer used.
  • Price: not published. Plans are scoped per organization; a free Context Impact Report runs on your own systems first.

Built for other jobs

Six products come up in the same searches and do a different job:

  • Zep builds dated context graphs, one per user or account, for customer-facing agents you build with its SDKs. It has the strongest model of time anywhere on this page. Its Context MCP server now opens a user's memory to off-the-shelf agents too; writes are on by default and an admin can make a connection read-only. See Harbor vs Zep.
  • Devin keeps its organization's knowledge as skills, after Knowledge notes were moved into a knowledge plugin per scope, and admins can have Devin change the organization's plugins, approving each change (release notes, 25 September 2026). It is memory for Devin, not for Claude Code, Cursor or Codex.
  • Gyld is a company brain for business operations: email, CRM, books and databases, reachable from coding agents over MCP but with nothing specific to code. $100 a month per workspace.
  • Glean is enterprise search across a company's apps, and its managed MCP server opens that index to Cursor, Claude Code, VS Code and Windsurf, with a plugin for Codex. Its engineering guide describes no review step for what gets indexed and nothing on what an agent used. It answers “where is that doc”, not “what did we decide”. Pricing is by demo.
  • Slite is a team wiki whose docs carry a verified state and a scheduled owner review, and its agent sends each change to the doc's owner first. Its MCP server makes the verified docs readable from Claude, ChatGPT and Cursor. It is the closest thing here to a gate on knowledge, kept for docs people write, not for what coding sessions learn. $10 or $20 a user a month.
  • Supermemory's company brain was open-sourced in September 2026 as a Slack teammate you host on Cloudflare. Its own guide says the self-hosted build has no MCP server or API yet, so a coding agent cannot read it.

Newer entrants that call themselves a company brain for agents, including Hyper, Memory Store and Almanac, launched this year. We have not reviewed them here yet.

Side by side

Rules come fromWho decidesReaches the agentShows what was usedBlocks a command
CLAUDE.md and AGENTS.mdWhoever writes themWhoever commitsThe whole file, every sessionNoNo
HarborSlack, PRs, docs and sessionsA gate you setRules that apply, at session startServed and cited, per ruleYes
UnblockedAbout 45 sourcesAutomaticWhen the agent calls a toolTool calls per agentNot documented
PackmindA playbook you writeDirect edits, or review on EnterpriseRule files in the repoWhich version is whereNo, but a pre-commit linter on Enterprise
CogneeSessions, plus 3 sources in CloudAutomaticEvery prompt, through hooksCitations, coverage scoresNo
ByteRoverFiles you uploadNo review step documentedWhen the agent calls its MCP serverFile and page on each passageNot documented
Mem0's pluginSession activityAutomaticUp to five memories on the first promptWhat was added or servedNot documented
MemoryLakeWhat AI tools write, plus importsNo review step describedMCP or post-commit hooksHistory and diffs of memoryNot documented
Sentra200+ company systemsA person approves what it sends, not what it storesWhen the agent calls its MCP serverNot documentedNot documented
Coworker50+ connectorsNot documentedWhen the agent calls its MCP serverNot documentedNot documented

Read the “Who decides” column again. A month ago a review step was rare. Now Packmind and Devin each have one, and Zep can make a connection read-only. A gate is becoming table stakes. The column that still separates these tools is “Shows what was used”: whether you can see, per rule, what reached a session and what an answer cited. A cite is evidence the rule was used, not proof it was obeyed.

How to choose

  • Your rules fit in one file and one person keeps it: stay with CLAUDE.md and AGENTS.md.
  • You want one playbook, versioned and reviewed, in every agent's own format: Packmind.
  • You want answers on demand across every system, with SSO and SOC 2 today: Unblocked.
  • You want to host the memory yourself, under an open licence: Cognee.
  • You want agents to look things up in the documents you upload: ByteRover. The fastest automatic memory: Mem0's plugin.
  • You want repo memory that branches and merges with your pull requests: MemoryLake.
  • You want one memory of the whole company, from sales to finance, behind a demo and an enterprise contract: Sentra or Coworker.
  • You want the decisions from Slack and reviews in every session, with a count of which rules were served and cited, and a gate you set: Harbor. Sign up, name your workspace, then run npx harborloop init.

Every comparison we keep is on one page, each dated and sourced to the other product's own site. If a line here is out of date, tell us and we will fix it.

Questions

What is the best company brain for coding agents?

It depends on where your decisions live and what you need to see. For decisions made in Slack threads and pull request reviews, served to every session, with a count per rule of how often it was served and cited, Harbor. For answers on demand across many systems with enterprise controls, Unblocked. For one versioned playbook written into every agent's rule files, Packmind. For open-source memory you host, Cognee.

Is a CLAUDE.md or AGENTS.md file enough?

For a small team with rules one person can keep, often yes. They are free and live in git. They stop scaling when nobody has time to write every decision down, when every agent gets the whole file whatever the task, and when nothing notices that a line has stopped being true.

What is the difference between agent memory and a company brain?

Agent memory keeps what an agent noticed, usually for one person or one repo. A company brain holds what the team decided, from the places it decides, and decides who approves a fact before every agent on the team is served it.

Which company brain tools work with Claude Code, Cursor and Codex?

The first six on this list reach Claude Code, Cursor and Codex. Harbor wires those three and GitHub Copilot with hooks and an MCP server. Unblocked, Cognee and Mem0 name more agents, ByteRover connects any agent that supports MCP, and Packmind writes rule files for ten agent formats. MemoryLake names Claude Code and Cursor, and Sentra and Coworker reach them through an MCP server the agent calls.

Are there open-source company brain tools?

Cognee and Packmind's core are open source under Apache 2.0 and can be self-hosted. Mem0's memory library is Apache 2.0, though its coding plugin needs the hosted platform. Harbor's CLI is MIT licensed and the service is hosted.

More on Memory tools and alternatives

  • Unblocked alternatives for coding agents (2026)
  • claude-mem alternatives for Codex, Cursor and teams
  • Supermemory alternatives for coding agents (2026)
  • Mem0 alternatives for coding agents (2026)

Harbor compared with each tool →

Your team’s rules, in every agent, checked on every PR.

Harbor is shared memory for your team’s Claude Code, Cursor, Codex and Copilot. Free for you and one teammate. No card. No end date.

Start free →

Product

  • How it learns
  • Only what fits
  • Pull request review
  • Keep it true
  • Handoffs
  • The result
  • Security
  • For coding agents
  • Pricing

Developers

  • Docs
  • Quickstart
  • Install
  • CLI
  • MCP server
  • Served and cited
  • Rule-file checker
  • Blog

Guides

  • Why Claude ignores CLAUDE.md
  • AGENTS.md vs CLAUDE.md
  • Share Claude Code memory with a team
  • Why Codex ignores AGENTS.md
  • Why Cursor ignores rules
  • Copilot instructions vs AGENTS.md
  • Shared memory across agents
  • Copilot Memory and Dreaming

Compare

  • All comparisons
  • vs Mem0
  • vs Zep
  • vs Unblocked
  • vs ByteRover
  • vs Cognee
  • vs Hindsight
  • vs OpenViking
  • vs Supermemory
  • vs Glen
  • vs MemPalace
  • vs Sentra
  • vs Trail-code

Company

  • Questions
  • Contact
  • X
  • LinkedIn
  • Privacy Policy
  • Terms of Service
  • Data Processing Agreement
  • Refund Policy
© 2026 HarborConnect your agents