Agent memory needs its OpenTelemetry moment

Agent memory is fragmented across repository instruction files, model-managed notes, hosted facts, context blocks, extracted memories, key-value stores, and knowledge graphs. Each format makes different choices about what counts as a memory, where it applies, how it is identified, and when it expires, so a lesson captured by one tool may be inaccessible to another. The central argument is that the problem is not a shortage of memory products but the absence of shared conventions. OpenTelemetry offers a precedent: competing observability projects joined around a common data model, naming conventions, protocol, and implementations, while vendors continued to compete on storage and analysis.

Existing memory proposals mainly address exporting a person’s data between platforms. Coding agents also need live shared addressing, so CI jobs, editors, and review bots can read and update the same repository-level memories. A common specification could define each record’s address, stable identity, content, lifecycle, and writer, without prescribing how memories are stored or ranked. It could also standardize names and operations, use MCP for access, and test that records and unknown metadata survive transfers. Security matters because shared memories can carry prompt injections; recording authorship, supporting signatures, and linking facts to relevant code could help readers assess trust and freshness. Key unresolved choices include address levels, default expiration, code references, and who should govern the work. The proposal calls for existing tools and standards efforts to converge on a neutral agreement rather than introduce another isolated format.