AI Knowledge Platforms for Storing Knowledge and Running Agents
Decide whether to keep agent knowledge in a separate store or inside an agent runtime, and see how AKB fits as the knowledge layer next to a runtime.
Choose the knowledge layer and the runtime separately
- Keep knowledge in its own layer when multiple agents, clients, or teams need shared access.
- Use AKB as that layer: documents, tables, files, explicit relations, and revision history for documents.
- Run agents where they already operate: Claude Code, Cursor, Codex, or another runtime can connect to the same authorized store.
- Add curation through integrations: external agents can process write events; AKB does not provide built-in curation agents or a scheduler.
AI knowledge platforms include dedicated knowledge stores, agent-execution runtimes, and products that combine both. Choose those responsibilities separately: AKB supplies the knowledge layer, while your chosen runtime handles agent execution, planning, tool calls, and scheduling.
Why separate the knowledge layer from the runtime?
Memory tied to one runtime may be available only to agents in that runtime. When shared access is unavailable, another tool may require a separate copy or an integration to use the same information. A separate knowledge layer gives authorized agents one shared place to read and write.
This separation also lets teams change runtimes or add scheduled agents without moving the knowledge. Document revision history remains available in the knowledge store.
| Layer | Job | Examples |
|---|---|---|
| Knowledge layer | Store, version, search, and permission organizational knowledge | AKB |
| Agent runtime | Plan steps, call tools, manage sessions; scheduling depends on the runtime | Claude Code, Cursor, Codex, custom runtimes |
| Curation | Summarize, clean up, or digest stored knowledge | External agents subscribed to write events |
What does AKB store, and how do agents use it?
AKB stores documents, tables, and files in vaults, which organize knowledge and access. Documents use native revisions in PostgreSQL. akb_history lists document revisions; use akb_get with a revision to read a previous version. Existing installations on the explicit Bare Git backend remain supported.
Agents work through MCP tools to:
- Write and edit documents:
akb_put,akb_update,akb_edit - Find information:
akb_searchfor hybrid semantic and keyword retrieval,akb_grepfor literal matches,akb_browsefor the collection tree, andakb_drill_downfor section-level retrieval - Connect knowledge:
akb_link,akb_relations, andakb_graphfor explicit relations such asdepends_on,related_to, andimplements - Control access:
akb_grant,akb_revoke, andakb_set_publicfor user, organization, or public access
Hybrid search, literal matching, and section-level retrieval help agents find relevant information in different ways.
How do runtimes connect?
MCP clients connect to AKB using the setup appropriate to each client. For current client options and setup instructions, see the AKB repository.
Each user connects with their own credentials and can access only authorized vaults. A shared AKB server does not automatically share a user's private session memory.
AKB plugins support specific workflows: akb-wiki ingests sources and answers questions from a vault with citations; akb-sessions turns a coding session into structured notes; and akb-claude-code uses Claude Code lifecycle hooks to load preferences and recent learnings, then write a recap. Session capture depends on the client integration and its configuration; akb-claude-code is specific to Claude Code.
For example, a coding assistant can record a project decision in a shared vault. Another authorized agent can reuse that decision and retrieve the document revision that contains it.
Where does curation happen?
AKB is a read/write store, not a built-in curation agent or scheduler. Writes are recorded in a PostgreSQL outbox; when Redis is configured, events can be streamed to external consumers.
Teams may connect integrations such as a weekly-digest agent or a stale-document checker. These are optional external consumers, so curation can change independently of the knowledge store.
When is a single all-in-one runtime enough?
One runtime with built-in memory may be enough when one agent serves one workflow and no other tool or team needs that memory. Consider a separate knowledge layer when another agent, tool, or team needs shared access, or when document revision history and vault-level permissions matter.
FAQ
Does AKB run agents?
No. AKB stores and serves knowledge. Agents run in Claude Code, Cursor, Codex, or another runtime and connect to AKB over MCP.
Can agents write to AKB as well as read?
Yes. Agents can create and edit documents through MCP tools, and AKB keeps document revision history.
Learn More
See the MCP tools, plugins, and setup guide in the AKB repository.
