AI brains landscape · snapshot June 2026

AI brain providers, compared.

A working map of where AI-brain tools sit in the stack: memory APIs, company knowledge graphs, retrieval plumbing, workflow runtimes, app/workspace layers, and enterprise assistants. Start with the use-case map, then use the table.

Use-case categories

Where each tool sits in the stack

“Open source” and “SaaS” are buying attributes, not the main category. The more useful question is: what job does this thing do? Click any tool below for details.

Provider	Use-case category	Principal use case	Tags	Open?	SaaS?	Pricing	Memory model	Freshness

Source note: open-source dates, licenses, and update signals use GitHub metadata checked June 19, 2026. SaaS pricing is summarized from public pricing pages where available and should be rechecked before buying or publishing procurement guidance.

Thesis: the useful split is shared company memory (“what does the organization know?”) versus agent-specific adaptive memory (“what has this AI worker learned from doing the job?”). Serious systems eventually need both, plus governance.

Evaluation framework

Canonical categories and criteria

The filters above separate use-case categories from buying attributes like SaaS and open source. The list below is the rubric I’d use for deeper scoring.

Memory API Company brain Retrieval plumbing Workflow/runtime Workspace/app Enterprise assistant SaaS Open source

1. Center of gravityMemory API, company graph, RAG framework, workflow runtime, enterprise search, or app platform.

2. OpennessOpen-source license, source-available, hosted-only SaaS, self-hostable, or enterprise deployment.

3. Pricing modelFree OSS, hosted starter plans, usage-based, seats, enterprise contract, or unclear/contact sales.

4. Memory modelUser memory, agent memory, session memory, entity memory, temporal events, graph memory, or document context.

5. Context ingestionConnectors, APIs, file ingestion, chat history, CRM/support/docs/email, webhooks, and agent traces.

6. Retrieval modelVector, keyword, hybrid, graph traversal, temporal search, reranking, ontology-aware retrieval, or context assembly.

7. Skill / workflow contextCan it store playbooks, instructions, tool recipes, skills, procedures, and reusable agent behavior?

8. Learning loopCan agents write back observations, outcomes, corrections, eval failures, and improved workflows safely?

9. GovernancePermissions, tenant isolation, provenance, audit logs, retention/deletion, approvals, and memory poisoning defenses.

10. Human inspectabilityCan humans see, edit, challenge, delete, and correct memory instead of treating it as fog in a jar?

11. Integration surfaceSDKs, REST API, MCP/tools/plugins, web app, CLI, cloud sync, and compatibility with existing agent runtimes.

12. Freshness evidenceLaunch/created date, latest release or push, changelog activity, docs updates, and product momentum.

Current thesis

The winner probably needs two kinds of memory.

Shared company memory answers: what does the organization know across docs, customers, workflows, decisions, and tools?

Agent-specific adaptive memory answers: what has this AI worker learned from doing its job for this customer, team, or process?

The serious product will need both, plus governance. Otherwise you either get a smart search box that never learns from work, or a learning agent with no trustworthy company context.