Context Repositories: Git-based Memory for Coding Agents — Letta

Summary

Letta Code’s new approach to agent memory: store context as local filesystem files backed by git. Agents manage their own context programmatically — reorganizing files, updating frontmatter, controlling what’s pinned to the system prompt. Git enables concurrent multi-agent memory processing.

Key Claims

  1. Files as universal memory primitive — both humans and agents can work with files using standard tools (bash, scripts, editors)
  2. Progressive disclosure via filesystem — directory structure + file frontmatter act as navigational signals. system/ dir = always in prompt.
  3. Git enables concurrent memory — multiple subagents can process/write memory in isolated worktrees, then merge. Solves the “single-threaded learning” problem.
  4. Memory skills: initialization (bootstrap from codebase + chat history), reflection (sleep-time background updates), defragmentation (reorganize messy memory)
  5. Agents manage their own context — not just read/write, but reorganize, split, merge files to optimize their own memory

Why It Matters

Direct evolution of MemGPT. The git-backed approach solves versioning, collaboration, and concurrent processing in ways that tool-based memory can’t. The “agents managing their own context” pattern is relevant to how we do MEMORY.md — we could adopt similar progressive disclosure.

Connections