Every agent framework reinvents memory from scratch. LangChain has its memory classes. CrewAI has its knowledge system. AutoGen has its teachability module. None of them talk to each other. Switch frameworks and your agents start from zero.

This is the same mistake the industry made with tool calling before MCP standardized it.

Today we’re releasing A2M (Agent-to-Memory) — an open protocol that standardizes how AI agents store, retrieve, validate, and reason about memory. And Seidr, a Tier 3 cognitive implementation that does the heavy lifting.

What A2M defines

A2M is a JSON-RPC 2.0 protocol with six methods:

  • memory/store — persist memories with type, agent attribution, and metadata
  • memory/query — semantic search with similarity scoring
  • memory/validate — confirm, contradict, or mark memories irrelevant based on task outcomes
  • memory/reflect — cluster and summarize related memories
  • memory/lifecycle — prune, delete, expire, consolidate, and promote memories
  • memory/status — health and statistics

Three capability tiers let backends adopt incrementally. Tier 1 is store-and-retrieve — you can build a conformant backend in 234 lines of Python. Tier 3 adds cognitive processing: fact extraction, relationship mapping, confidence scoring, and cross-agent knowledge sharing.

Discovery works through a capability card at /.well-known/a2m.json. An agent runtime reads the card, learns what the backend supports, and adjusts its behavior. No configuration files, no environment variables — just a URL.

Why this matters

Agent memory isn’t a nice-to-have. It’s the difference between an agent that makes the same mistakes repeatedly and one that compounds knowledge across tasks.

When our agent elder reviews a PR and discovers that NetworkPolicy egress rules need explicit DNS allowances, that lesson should persist. The next time any agent in the warband writes a NetworkPolicy, that memory should surface automatically. A2M makes that workflow possible — and portable.

The protocol is also designed for sovereignty. Memories live in your database. The provider runs on your infrastructure. There’s no cloud dependency, no vendor lock-in, and a clear audit trail for every memory operation.

Seidr: the reference implementation

Seidr is a Tier 3 cognitive A2M backend. It runs on PostgreSQL with pgvector, uses GPU-accelerated embeddings, and exposes both the A2M JSON-RPC endpoint and an MCP server for direct agent integration.

When Seidr receives content above a configurable threshold, it runs a cognitive pipeline: extract atomic facts, map relationships between entities, and assign confidence scores. Over time, validation signals from agent task outcomes adjust these scores — memories that lead to successful tasks gain confidence, memories that lead to failures get flagged.

Seidr is open source under MIT and runs anywhere PostgreSQL runs.

Getting started

The fastest path is the Getting Started guide. If you want to explore the protocol interactively, try the Protocol Explorer.

The spec, a Tier 1 reference implementation (tiny-memory), and Seidr are all on GitHub.

A Sovereignty Labs project. Own your intelligence.