SAGE (short for Sovereign Agent Governed Experience) is a persistent project memory system for AI agents. Its core goal is to solve the problem of "having to re-explain the project to the AI every time": it lets agents record project decisions, approaches that were tried and failed, and useful lessons learned, and recall relevant context in subsequent sessions. SAGE supports Claude Code, Codex, Cursor, and other MCP (Model Context Protocol)-capable clients, and developers can also integrate via the Python SDK or a signed REST API.
Main capabilities include:
- Memory recording and recall: agents can record experiences and retrieve relevant memories in subsequent sessions. Recall effectiveness depends on what has been recorded, the configuration, and how the agent uses the tools.
- Provenance and admission: agent-signed writes attribute their records to the corresponding author; memories must go through a verification lifecycle before entering normal recall.
- Access control: registration, domain permissions, and classification determine which agents can use a given memory.
- Memory maintenance: you can inspect confidence and status, apply decay, or deprecate a memory while preserving audit history.
- CEREBRUM dashboard: a browser-based operations console centered on a 3D MRI memory brain, including features such as chain health, agents, federation, semantic memory, recall tuning, vault recovery, tasks, import, and updates.
How to use:
- Desktop: download the latest version, open SAGE, and complete setup in CEREBRUM. Choose a memory engine, connect AI clients, and review agent access permissions.
- Build from source: requires Go 1.26.9+.
- Container image: ghcr.io/l33tdawg/sage:11.23.23.
- After connecting an MCP client, call sage_inception and follow the memory schema it returns.
- HTTP MCP is also available at /v1/mcp/sse and /v1/mcp/streamable, but both require bearer token or OAuth authentication.http://localhost:8080 is the REST base address, not an unauthenticated MCP endpoint.
Applicable scenarios:
- An agent needs to participate continuously in the long-term work of the same project.
- Valuable design decisions, failed approaches, or lessons for the next session need to be preserved.
- Unlike RAG, SAGE provides, beyond retrieval, a governed record around agent experience.
Limitations and cautions:
- Memory is stored locally by default. Optional federation, connectors, and remote model providers may send data to configured destinations. Enabling the vault encrypts memory content at rest.
- Validator checks govern what is submitted, do not prove that every memory claim is true, and do not change model weights.
- Personal mode runs a real CometBFT validator with no Byzantine redundancy; registering more agents does not add consensus validators. Multi-validator deployments run a shared chain; federation connects different chains under explicit policy.
- Agents are not validators, and their requests and results remain untrusted data.
- If upgrading from a version installed before v5.0 and AI has not performed per-turn memory updates, rerun the installer in the project directory.
- Before version upgrades, read the corresponding guide; multi-admin chains especially require this before applying the v23 activation.
- Some features (such as the local judge memory gate) remain experimental and are off by default.
New in the current version v11.23.23: close a reviewed notification without replying via sage_message_close(message_id, reason), record a "closed but not replied" outcome, preserve history, and clear pending reminders. Application version v28 remains active, with no chain migration required.
