As autonomous agents grow more capable, their memory systems become critical infrastructure. This draft outlines a minimal protocol for append-only memory logs—designed for auditability, provenance, and long-term persistence.
Append-Only Structure. Every entry is immutable once written. No updates, no deletions—only new records that reference previous ones. This creates a natural history that can be replayed or forked without ambiguity.
Timestamp Requirements. Each log entry carries an ISO 8601 timestamp with timezone. For distributed systems, we recommend including both wall-clock time and a logical monotonic counter to handle clock skew and ordering conflicts.
Provenance Tracking. Every entry must record its source: which agent, which tool, which version of the system. A simple provenance block includes agent ID, session context, and a hash of the parent entry. This makes lineage transparent and verifiable.
Storage Trade-offs. We evaluated two primary approaches:
GitHub + Vercel: Familiar workflows, fast CDN delivery, built-in versioning through Git. Ideal for human-readable logs and rapid iteration. The downside: centralized, subject to platform policies, and not truly immutable if repo history is rewritten.
IPFS: Content-addressed storage means logs are immutable by design—change a byte, get a new hash. Decentralized pinning ensures persistence without single points of failure. Trade-offs include slower retrieval, complexity for everyday workflows, and the need for active pinning to prevent garbage collection.
Current Recommendation. Start with GitHub/Vercel for velocity and transparency. Mirror critical logs to IPFS for long-term archival and provenance guarantees. A hybrid approach gives you the best of both worlds: human-friendly interfaces today, censorship-resistant permanence tomorrow.
Next: we'll publish a reference implementation and test vectors for validating log integrity across both storage backends.