Eazler AI Lab AI Engineering Journal

Agent memory — long-term memory structures and file-based approaches

Agent memory is the question of where, and in what format, to keep information that must outlive a conversation. The context window disappears when a session ends and cannot grow without limit. Until the first half of 2026 the closest thing to a standard answer was vector databases and knowledge graphs; this year, keeping memory in a single folder of Markdown files has taken hold. People can read and edit it, git keeps its history, and search ends with grep.

3 reports55 min totalUpdated 2026-10-06

What these reports establish

The path from Karpathy's LLM wiki gist to Google Cloud's Open Knowledge Format v0.1 and to v0.2, which adds trust signals; the line between what the specification actually defines (one required field, type) and what tools and the community claim; and the Brain Layer architecture that ties memory, search, citation and knowledge-graph reasoning together.

Memory: 3 reports

All reports →