Emem: A Signed Memory Layer for Multi-Agent AI Systems
Emem gives AI agents a signed, verifiable memory layer for physical-world facts that survives context compaction across models and vendors.
Emem is a new memory layer designed to solve a specific problem in multi-agent AI systems: when an agent's context window is compacted, verified facts often degrade into imprecise paraphrases that drift further from the truth with each summarization pass. Emem addresses this by storing facts as small, cryptographically signed records at permanent, content-derived addresses, starting with Earth observation data such as elevation measurements.
Instead of retrieving documents from a trusted store, agents exchange short 'fact tokens' (e.g., emem:fact:...) that resolve to the exact same signed bytes on any model, any vendor, at any later time. Any agent can verify a fact's authenticity offline using ed25519 signatures, without trusting the sender or the server hosting emem. This makes it possible for agents from different companies, with no shared infrastructure, to agree on the same physical-world fact and independently verify it.
The system is accessible via MCP integration, a REST API, and Python/TypeScript SDKs, with no key required for reads. The team is explicit about trade-offs: fact tokens cost roughly 5.8x more context than pasting a raw number, so the approach is best suited for long-running tasks, cross-trust verification, audit trails, and multi-robot fleet coordination — not for conversational memory or high-frequency data streams.
For engineers building multi-agent architectures, this is a concrete example of designing memory that survives context compaction and crosses trust boundaries, rather than relying on RAG-style retrieval from a single vendor's store.