Zaxy

About Zaxy

Zaxy is an MIT-licensed memory substrate for fleets of AI agents, built around an append-only, hash-chained event log it calls Eventloom. Everything the system exposes — recall, rules, consolidated knowledge — is a projection of that log, so deleting a projection is safe: replay rebuilds it, and nothing authoritative lives anywhere else. Each recall is a Memory Checkout that returns cited results with a citation URI pointing back to the exact event, and each change is itself a hash-sealed event, which makes the whole memory replayable, rollback-able, and auditable. The governance model is the distinguishing feature. Memory only changes through an evolution gate configurable as auto, propose, or review, so an agent cannot silently rewrite what the fleet believes. An outcome loop lets agents report success or failure and turns lessons into cited preventive rules; idle-time crystallization merges near-duplicates without runtime cost; a fleet memory plane propagates rules across agents with trust tiers and visibility scopes; human edits and rollbacks are reversible cited events that leave originals intact; and verified forgetting uses crypto-erasure to destroy a payload key while the chain still verifies. It installs via a shell script that wires up every agent harness it detects, or via uvx, pip, or MCP config, and ships 49 MCP tools. The vendor publishes benchmark numbers with unusual candour, reporting LongMemEval-S at 0.90 with a gpt-5 reader on the full 500 and explicitly retracting prior oracle-mode claims.

Key Features

Zaxy Pros & Cons

Who Is Zaxy Best For?

Tags