A persistent, BFT consensus-validated memory infrastructure for AI Agents, ensuring integrity and consistency of every memory through a CometBFT consensus engine and quadruple application validators.
SAGE is a persistent memory infrastructure designed for AI Agents, with its core innovation being the introduction of Byzantine Fault Tolerant (BFT) consensus mechanisms from distributed ledgers into the Agent memory layer. Every memory write must pass a 3/4 BFT vote across four application validators — Sentinel, Dedup, Quality, and Consistency — implemented on CometBFT ABCI 2.0 with consensus primitives equivalent to distributed ledgers.
The project offers two operating modes: Personal (single node) and Quorum (multi-node network). Personal mode runs a real CometBFT node with 4 in-process validators; Quorum mode supports multi-node deployment with TLS-encrypted communication (v6.5 adds automatic ECDSA P-256 certificate authority) and a dual-listener architecture.
On the security front, all memory writes are protected by Ed25519 digital signatures. The storage layer uses AES-256-GCM encryption with Argon2id key derivation (Synaptic Ledger), supporting vault lock/unlock. Inter-node communication in Quorum mode is encrypted via TLS, while the CometBFT P2P layer uses SecretConnection (X25519 DH + ChaCha20-Poly1305).
For memory management, each memory carries a confidence score that naturally decays over time. FTS5 full-text search complements semantic embeddings, and two memory modes (full/bookend) are available to optimize token usage. v6.0 introduced dynamic validator governance, supporting zero-downtime addition/removal of validators through on-chain governance proposals with 2/3 BFT voting, including safety constraints such as a maximum 1/3 power limit.
For external interfaces, SAGE connects to AI systems including Claude, ChatGPT, DeepSeek, and Gemini via MCP protocol and REST API, with a Python SDK available (v6.1.0+). The CEREBRUM dashboard provides force-directed neural graph visualization, domain filtering, semantic search, and real-time SSE updates.
Empirical validation through a 50-vs-50 controlled experiment on the Level Up CTF platform demonstrated significant advantages for agents with memory. Longitudinal research showed that agents with SAGE memory exhibited cumulative learning trends (Spearman rho=0.716, p=0.020), while the control group showed no learning trend (rho=0.040, p=0.901). Performance benchmarks reached 956 req/s commit throughput and 21.6ms P95 query latency.
The codebase is written in Go (62%), JavaScript (21%), and Python (11%), currently at v6.5.5, supporting macOS / Windows / Linux cross-platform operation and Docker deployment. Code is under Apache 2.0, papers under CC BY 4.0. Created by @l33tdawg (Dhillon Andrew Kannabhiran) with 4 Zenodo papers (@misc type, peer review status unconfirmed).