Browse the
library.

Interleaved reasoning traces and actions — the canonical paper on agents that think and act in a loop.

LLMs as OS-like memory managers: paged context, self-editing memory. Foundation of the Letta stack.

The benchmark that made coding agents measurable — real issues, real patches, verified.

Deliberate search over thought trees — planning as exploration, not a single chain.

Programming, not prompting: compiled, optimizable pipelines with typed behavior.

Hippocampus-inspired indexing for long-term agent memory — knowledge graphs meet retrieval.

Verbal self-reflection as memory — agents that learn from their own failures.

The original RAG paper — parametric and non-parametric memory, combined.

The standard interface for tools and context: the connective tissue Alexandria's machine surface speaks.

Stateful agents with OS-like memory management — the productionized line from the MemGPT paper.

Extracted, consolidated agent memory as a drop-in layer.

The reference terminal-native coding agent: git-native, editor-optional.

The agent-computer interface design that made real GitHub issues tractable.

Canonical recipes for tool use, structured output, and agent patterns.