RC RANDOM CHAOS

Standardizing Agent Memory Formats to End Tool Fragmentation

· via Hacker News

Original source

Agent memory needs its OpenTelemetry moment

Hacker News →

Agents and development tools store memory in incompatible formats: instruction files, API calls, GitHub-stored facts, MCP knowledge graphs. A lesson learned by one agent—that tests need PostgreSQL running, for instance—stays trapped in its format. An editor cannot read what a CI agent discovered. Each new memory tool solves the same problem independently, multiplying formats without reducing fragmentation.

OpenTelemetry unified observability by getting competing vendors to agree on what not to compete over: a shared data model, semantic conventions, a protocol, and reference implementations. Vendors continued competing on storage, querying, and analysis. Agent memory stands at that pre-unification stage. The Linux Foundation’s Agentic AI Foundation already hosts MCP and AGENTS.md standards; a shared memory spec could live there. Existing proposals show what works: deletion tombstones, minimal cores, proper time modeling.

A complete spec must establish where memory applies (repo, branch, global), how identity works, what expires when, and who wrote it—enabling readers to judge trustworthiness. It should leave competition to features: search ranking, embeddings, storage, user interfaces. The hard problems remain unresolved: how agents coordinate writes to shared memory, and how to make injection vectors visible without blocking legitimate sharing.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.