What is persistent memory for an agent?
Persistent memory is information an AI agent can retrieve after an earlier conversation has ended. It may include facts, decisions, procedures, preferences, or documents that remain useful across tasks.
The key word is persistent. A context window is the text an AI can consider during one interaction. It is not durable storage. Persistent memory requires an external system where information is saved, organized, searched, and returned when needed.
- Context window: temporary working space for one interaction.
- Pasted notes: exact context, but manually repeated.
- Uploaded files: useful source bundles for a specific task.
- RAG: retrieval from a broader document collection.
- Curated memory: maintained facts, decisions, and notes.
How does context window memory work?
A context window contains the system instructions, conversation messages, retrieved material, and task data currently available to the model. Longer inputs can help with complex work, but they do not automatically create lasting memory.
As conversations grow, relevant details may be crowded out or summarized. Context-window limits and product-specific behavior also vary. For reliable recall, agents should retrieve only the information needed for the current task rather than expect every past detail to remain available.
- Identify the facts or documents relevant to the task.
- Retrieve only the necessary passages.
- Place them in the current context with clear source labels.
- Save durable changes back to an external memory system when authorized.
What is RAG for AI memory?
RAG stands for retrieval-augmented generation. Before answering, an AI system searches a collection of documents, retrieves relevant passages, and includes them in the model’s current context. The answer is then grounded in retrieved material rather than model recall alone.
RAG works well for large, changing collections because it retrieves current content instead of training the model again. It also needs good source quality, useful organization, and careful retrieval. Irrelevant passages can weaken an answer, and an answer without citations still requires verification.
- Broad document sets
- Frequently updated sources
- Source-grounded answers
- Need for retrieval relevance testing
How does a curated knowledge base compare?
A curated knowledge base emphasizes material you deliberately save, organize, and maintain. It can hold full source text alongside highlights, notes, decisions, and playbooks. This creates a more controlled memory than blindly retrieving every related document.
KnowMan uses this approach. It reads the full text of saved items and organizes articles automatically. Its MCP server exposes six tools: kb_search, kb_get_item, kb_ask, kb_memories, kb_recent, and kb_save. Compatible agents can search, read, ask with citations, consult memories, view recent saves, and save links.
KnowMan supports Claude Desktop, Claude Code, Codex, Cursor, and any MCP client. Agents can save items but cannot write memories, so your decisions and playbooks remain under your control.
- Save authoritative articles and reference material.
- Add private highlights, notes, and decisions where useful.
- Connect an MCP client using an agent key with chosen permissions.
- Ask questions that require search, full-text reading, or citations.
Which memory approach should you choose?
Choose pasted notes or uploaded files for a small, immediate task. Choose RAG when you need to search a large and changing document collection. Choose a curated knowledge base when you value durable context, source traceability, and control over what an agent treats as memory.
Hybrid setups are common. An agent may use chat context for the current task, retrieve documents for evidence, and consult saved memories for standing decisions. No approach removes the need to inspect important answers.
- Quick, one-off work: paste or upload.
- Large changing collections: RAG.
- Durable personal context: curated knowledge base.
- Sensitive information: review permissions and saved content.
Use the hosted KnowMan server
Connect your client to the hosted KnowMan server at https://api.knowman.io/mcp. Add the server URL in an MCP client with OAuth support, then sign in on the KnowMan consent page and approve access. KnowMan access tokens last one hour and refresh automatically.
You can also create an API key in KnowMan settings and send it as an Authorization: Bearer <key> header. For Claude Code, run claude mcp add --transport http knowman https://api.knowman.io/mcp --header "Authorization: Bearer <key>". For Cursor, use {"mcpServers":{"knowman":{"url":"https://api.knowman.io/mcp","headers":{"Authorization":"Bearer <key>"}}}}.
The open-source client is available at https://github.com/Essatt/knowman-mcp. Once connected, ask the assistant to search your KnowMan knowledge base or save a link.
Questions people ask
Is a context window the same as long-term memory?
No. A context window is temporary working space for the current interaction. Long-term memory requires information to be stored outside the conversation and retrieved later.
What is the difference between RAG and a personal knowledge base?
RAG is a retrieval technique that can use any document collection. A personal knowledge base is the curated collection itself, containing the sources, notes, and decisions you choose to maintain.
Can an AI agent save new information to its memory?
That depends on the system and its permissions. KnowMan agents can save links with kb_save, but they cannot write memories; decisions, playbooks, facts, and notes remain yours.
Which clients can use the KnowMan MCP server?
KnowMan works with Claude Desktop, Claude Code, Codex, Cursor, and any MCP client. Setup and AI assistant connection are Free.