Agent memory needs its OpenTelemetry moment

Today's coding tools keep separate and isolated records of everything they remember, each in their own unique format that cannot read or write to the others. These distinct shapes include CI agents, editor agents, and memory layers, all attempting to solve the same foundational issue. This situation echoes the challenges faced in observability before 2019, when each vendor provided their own agents, instrumentation libraries, and data formats.

Switching vendors necessitated reinventing instrumentation. OpenTelemetry's integration in 2019 resolved this by establishing a shared standard, allowing vendors to focus on how they utilize the data rather than competing on format alone.

The situation with agent memory is analogous to that of observability, with numerous developers building their own memory solutions, each with its own storage, schema, and MCP tools. While this demonstrates the problem's prevalence, each new solution simply adds another unreadable format. The core issue lies in the lack of community agreement, which can only emerge through collective effort.

LoreKit, an open-source tool created by the author, addresses this issue. However, it represents only one proposal among many, each with its own shortcomings. The objective is to establish a shared standard, similar to how OpenTelemetry emerged as the solution for observability fragmentation. By addressing the four fundamental questions—where the memory is true, its identity, its content, and its expiration—a shared spec can unite the fragmented ecosystem.

The record format proposed includes frontmatter for contextual information, such as address, key, identity, lifecycle, and authorship. The actual memory content follows, with content_type specifying how it should be interpreted. This structure mirrors OpenTelemetry's layered approach, separating concerns like storage, ranking, embeddings, and user interfaces, which should remain the domain of individual tools. The spec would be neutral, hosted in a neutral space rather than tied to a single vendor or tool.

While shared memory has potential as a prompt-injection channel, proper specification can mitigate risks such as poisoning. By recording authorship, allowing signatures, and linking to source code, a spec can ensure transparency and accountability. Standardization, like OpenTelemetry, can address fragmentation by merging existing proposals and tools, rather than introducing additional formats.

The author believes that by combining the strengths of current solutions, the community can achieve a unified standard for agent memory.