(S)AGE — Sovereign Agent Governed Experience

Persistent, consensus-validated memory infrastructure for AI agents.

SAGE gives AI agents institutional memory that persists across conversations, goes through BFT consensus validation, carries confidence scores, and decays naturally over time. Not a flat file. Not a vector DB bolted onto a chat app. Infrastructure — built on the same consensus primitives as distributed ledgers.

The architecture is described in Paper 1: Agent Memory Infrastructure.

Just want to install it? Download here — double-click, done. Works with any AI.

Architecture

Agent (Claude, ChatGPT, DeepSeek, Gemini, etc.)
  │ MCP / REST
  ▼
sage-gui
  ├── ABCI App (validation, confidence, decay, Ed25519 sigs)
  ├── App Validators (sentinel, dedup, quality, consistency — BFT 3/4 quorum)
  ├── CometBFT consensus (single-validator or multi-agent network)
  ├── SQLite + optional AES-256-GCM encryption
  ├── CEREBRUM Dashboard (SPA, real-time SSE)
  └── Network Agent Manager (add/remove agents, key rotation, LAN pairing)

Personal mode runs a real CometBFT node with 4 in-process application validators — every memory write goes through pre-validation, signed vote transactions, and BFT quorum before committing. Same consensus pipeline as multi-node deployments. Add more agents from the dashboard when you're ready.

Full deployment guide (multi-agent networks, RBAC, federation, monitoring): Architecture docs

CEREBRUM Dashboard

CEREBRUM — Neural network memory visualization

http://localhost:8080/ui/ — force-directed neural graph, domain filtering, semantic search, real-time updates via SSE.

Network Management

Network — Multi-agent management

Add agents, configure domain-level read/write permissions, manage clearance levels, rotate keys, download bundles — all from the dashboard.

Settings

What’s New in v5.2.2

Memory Type Guidance — sage_remember tool description now guides agents to use the fact type for infrastructure details (IPs, hostnames, SSH commands) instead of observation, preventing decay across provider boundaries.

v5.2.1

Self-Updater Fix — Dashboard "Update" button no longer fails. GitHub changed release asset downloads to route through release-assets.githubusercontent.com, which wasn’t in the redirect allowlist.

v5.2.0

Immutable RegisteredName — Agents now have a permanent RegisteredName (set once at registration) alongside the mutable display Name. Renames preserve the original on-chain identity for audit trail and provenance. Dashboard shows "Registered As" when names differ.
Lazy Backfill — Pre-v5.2.0 agents automatically get their RegisteredName populated from their current name on first access.

v5.1.0

Agent Rename Fix — Renaming an agent in the dashboard (Network → Agents → Edit) now reliably syncs to on-chain state. Previously, the CometBFT broadcast was fire-and-forget and could silently fail, causing sage_inception to return the old auto-generated name.
Self-Healing Name Reconciliation — If on-chain and display names ever diverge, sage_inception automatically detects and repairs the mismatch on the agent’s next boot.
Broadcast Error Feedback — The dashboard now warns you if an on-chain sync fails instead of silently swallowing the error.

v5.0.12

MCP Identity Fix — sage-gui mcp and sage-gui mcp install --token now respect SAGE_IDENTITY_PATH environment variable as the highest priority (exactly matching the SDK’s AgentIdentity.default()).
Auto-creates directories + keypair if missing. Added clear INFO logs. 100% backward compatible.
Enables clean multi-agent setups (e.g. multiple Claude Code instances in tmux). Closes the identity collision issue.

What’s New in v5.0.11

Docker Fix — Container no longer stuck in restart loop. Default entrypoint changed from MCP stdio mode to serve (persistent REST API + dashboard). MCP stdio still available via docker run -i ghcr.io/l33tdawg/sage mcp. Fixes #14.

v5.0.10

Multi-Agent Identity — New SAGE_IDENTITY_PATH env var and AgentIdentity.default() for running multiple Claude Code agents on the same machine without key collisions. (Community PR by @emx)
Dashboard Fix — "Synaptic Ledger" label in overview settings now reads "Synaptic Ledger Encryption" to clarify it refers to the encryption state, not the ledger itself.

v5.0.9

Upgrade Hang Fix — Fixed CometBFT startup hang after drag-and-drop upgrades. Stale consensus WAL files left behind during migration caused a 60-second timeout and prevented the REST API from starting. Now cleaned up automatically at both migration and startup time.

v5.0.7

Agent Pipeline — Inter-agent message bus (sage_pipe) for direct agent-to-agent communication. Send messages, check results, coordinate work across agents in real-time.
Python Agent SDK — sage-agent-sdk on PyPI with full v5 API coverage for building SAGE-integrated agents. CI-tested on every release.
Vault Recovery — Reset your Synaptic Ledger passphrase using a recovery key. No more permanent lockouts.
Memory Modes — Choose full (every turn), bookend (boot + reflect only), or on-demand (zero automatic token usage) to control how much context your agent spends on memory.
Vault Key Protection — Vault key is automatically backed up on every upgrade and in-app update. Prevents the silent overwrite that could cause permanent memory loss.
macOS Tahoe Compatibility — Fixed Gatekeeper warnings and launch failures on macOS 15.x. Removed the Install SAGE.command that triggered quarantine blocks.
Linux ARM64 Containers — Docker images now build for linux/arm64 in addition to amd64.
/v1/mcp-config Endpoint — Agents can self-configure their MCP connection without manual setup.
Docker Images — Every release auto-builds and pushes to ghcr.io/l33tdawg/sage. Pin a version or pull latest.

v4.5

Cross-Agent Visibility Fixed — Org-based access (clearance levels, multi-org federation) now correctly grants visibility across agents. Queries and list operations check direct grants, org membership, and unregistered domain fallback — no more 0-result queries when clearance should allow access.
Domain Auto-Registration — First write to an unregistered domain auto-registers it with the submitting agent as owner and full access granted. No more propose-succeeds-but-query-404.
RBAC Gate Simplification — DomainAccess (explicit allowlist) and multi-org gates are alternatives, not stacked. Passing one skips the other.

v4.4

CEREBRUM UX Overhaul — Snap-back physics (nodes spring back to cloud on focus exit), forget animation (fade-and-remove instead of full reload), tab backgrounding fix (no physics jumps after alt-tab).
Clean Synaptic Ledger — Always-visible button with double-click confirmation. Cleanup toggle auto-saves.
Focus Mode — Single-click to view memory detail, side panel closes on exit. Graph defaults to committed status.

v4.3

Synaptic Ledger Safeguards — Three-layer defense against silent encryption downgrade: server auto-re-enables if vault.key exists, web login now actually unlocks the vault for writes (was a bug), and the native macOS app icon prompts for your passphrase before launch. Plaintext writes are blocked when the vault is locked.
Vault-Locked API — /v1/dashboard/health now exposes vault_locked status. MCP tools (sage_remember, sage_turn, sage_reflect) check this flag and return clear errors telling agents to prompt the user to unlock via CEREBRUM — no more silent plaintext fallback.
Isolated-by-Default RBAC — Agents can only see their own memories by default. Domain-level read/write permissions, clearance levels, and multi-org federation with department filtering.
Bulk Operations — Multi-select memories in CEREBRUM for bulk domain moves, tag additions, and agent reassignment.
Dashboard Update Check — Long-open tabs now poll for new releases every 12 hours so you never miss an update.
Automated Docker + MCP Registry — Release CI now auto-builds Docker images, pushes to GHCR, and updates server.json — MCP registries get new versions without manual intervention.

v4.0

4 Application Validators — Every memory now passes through 4 in-process validators before committing: Sentinel (baseline accept, ensures liveness), Dedup (rejects duplicate content by SHA-256 hash), Quality (rejects noise — greeting observations, short content, empty headers), Consistency (enforces confidence thresholds, required fields). Quorum requires 3/4 accept (BFT 2/3 threshold).
Pre-Validation Endpoint — POST /v1/memory/pre-validate dry-runs all 4 validators without submitting on-chain. Returns per-validator decisions and quorum result. MCP tools use this to reject low-quality memories before they hit the chain.
Memory Quality Gates — sage_turn filters low-value observations (greeting noise, short content). sage_reflect detects similar existing memories and skips duplic

Sage

Install / Use

README