The Agent Memory Mess: Why Your AI Tools Are Currently Amnesiacs

Let's be honest: most of the 'memory' in today's AI agents is a joke. As Mads Thines recently detailed via LoreKit, we are currently building expensive, unpredictable toys because every tool is speaking a different language. If your CI agent learns something critical about your codebase, that knowledge is trapped in a silo. Your editor agent can't access it, and the next tool you install certainly won't. We aren't building intelligence; we're building a digital Tower of Babel.

According to an analysis by Thines, the industry is currently plagued by a total lack of a shared data model for agent memory. Thines, who works with OpenTelemetry at Dash0 and develops the open-source LoreKit tool, points out that a typical developer setup might involve four or more incompatible memory shapes: `.mdc` rules in Cursor, `AGENTS.md` or `CLAUDE.md` files, Anthropic's memory tool writing to `/memories` directories, and GitHub Copilot storing repo facts on its own backend.

**Opinion:** We are essentially treating AI agents like fancy goldfish with five-second memories. Until we have a standardized observability layer, these tools will remain fragmented. We don't need more 'memory layers'—we need a missing agreement.

Thines draws a parallel to the observability crisis before 2019. Back then, every vendor shipped their own instrumentation and data formats, making switching costs astronomical. That ended when OpenTracing and OpenCensus merged into OpenTelemetry, establishing a shared data model, semantic conventions, a protocol (OTLP), and reference implementations. The vendors stopped competing on how data was sent and started competing on how it was analyzed. Agent memory is currently where observability was in 2018.

LoreKit's analysis reveals a chaotic landscape of existing formats as of late 2026:

* **Instruction files:** `AGENTS.md` and Cursor `.mdc` rules use Markdown files that never forget unless a human edits them. * **Agent-managed files:** Anthropic's memory tool uses files in `/memories` that only disappear if the model deletes them. * **Hosted facts:** GitHub Copilot uses code citations that expire after 28 days of inactivity. * **Knowledge graphs:** MCP reference memory servers use entities and relations that never forget. * **Temporal graphs:** Graphiti and Zep track facts on graph edges with separate timelines for when a fact was true versus when it was recorded. * **Namespaced stores:** LangGraph's BaseStore uses JSON value dicts with optional TTLs (time-to-live).

There are attempts to fix this, though none have achieved total dominance. LoreKit notes that MacPaw Research proposed 'Portable Memory' (a `.mem` bundle), while the Open Memory Protocol (an April 2026 draft) focuses on moving conversations via ZIP files. Additionally, Cognee's COGX format attempts to import from Letta, Zep, and Mem0.

However, Thines argues that most of these solve 'export'—moving data from Vendor A to Vendor B—rather than 'live addressing.' The real challenge is enabling a CI job, an editor agent, and a review bot to read and write to the same memory space in real-time. While the Linux Foundation's Agentic AI Foundation now houses both MCP and `AGENTS.md`, the industry is still waiting for its 'OpenTelemetry moment.'