Skip to main content

AI Agent Memory

Use Revisium as a structured memory layer for AI agents with typed schemas, version control, and human review.

When to Use​

  • AI agents that need persistent, structured memory (Claude Code, Cursor, custom)
  • Agent-generated data that needs human review before going to production
  • Multi-agent systems sharing state
  • Any scenario where agent-written data needs audit trail and rollback

Key Features​

  • MCP Protocol — built-in MCP server for direct agent access
  • Typed Schemas — agents work with structured data, not raw text
  • Version Control — rollback on corruption, audit trail of all agent changes
  • Human Review — Admin UI for reviewing what agents wrote before committing
  • Branching — isolate agent work per task or per agent

Setup​

# Start Revisium
npx @revisium/standalone@latest --auth

# Add to Claude Code
claude mcp add --transport http revisium http://localhost:9222/mcp

How Agents Use It​

The agent designs schemas, creates tables, and manages data through MCP tools:

Agent: "I need to store research findings"
→ login
→ createTable("findings", schema with title, source, content, tags)
→ createRow("findings", "finding-1", { title: "...", content: "...", ... })
→ commitRevision (with user permission)

Schema Design by Agents​

Agents can design their own schemas based on the task. For example, a research agent might create:

TableFields
topicsname, description
findingstitle, source, content, topic (FK → topics), relevance
decisionstitle, rationale, status, related_findings (FK array → findings)

The schema resource (revisium://specs/schema) is available to agents for understanding valid schema formats.

Human Review Workflow​

  1. Agent makes changes in the Draft revision
  2. Human opens Admin UI → Changes tab
  3. Reviews schema changes and data modifications
  4. Approves (commit) or rejects (revert)

This is a differentiator — no other AI memory solution provides visual review of agent-written data.

Branching for Agents​

BranchPurpose
masterReviewed, committed agent data
agent-researchAgent's working branch — human reviews before merging
agent-code-analysisAnother agent's isolated workspace

See Also​