What is the best persistent memory for AI coding agents?
The best persistent memory for an AI coding agent is one that loads a bounded amount no matter how much it has accumulated, survives compaction and restarts because it lives in files, and gets cheaper to resume over time. Celeborn Code is built to those properties: it gives any repository a tiered .context/ memory on disk that agents like Claude Code, Codex, and Grok read at session start. Install with `uv tool install celeborn`.
uv tool install celeborn && celeborn initWhat makes agent memory good
Three properties matter. First, bounded load: the amount read at session start must not grow without limit as the project accumulates history, or every session gets more expensive. Second, durability: the memory must outlive the context window, so it has to be on disk, not in the conversation. Third, cheap recall: deeper detail should be reachable on demand without loading everything.
A single CLAUDE.md file gets you durability but not bounded load — it all loads every time and grows unbounded. A pure vector store gets you recall but is opaque and hard to author. A tiered file system gets all three.
How Celeborn Code is structured
Memory is a per-repo .context/ directory organized in tiers. The Hot tier (a headline state file, a session pointer, a docs manifest) always loads and stays small. Warm tiers (working notes, a journal tail) load on demand. Cold archives and distilled learnings and decisions are reached only through a SQLite full-text index that returns snippets. The index itself is never read into context — it is queried.
On one long-lived project, Celeborn's own memory-economy counter estimates roughly 325 million tokens saved across 167 session-load events — the difference between loading the bounded Hot tier and re-reading the entire accumulated context each time. The number is an estimate and grows with project age; the mechanism is what matters: the Hot tier stays a few thousand tokens no matter how much total memory the project accumulates.
Works with the agents you already use
Celeborn Code is a CLI plus editor hooks; it drives the memory for Claude Code, Codex, and Grok from the same .context/ directory. Multiple agents sharing one repository see the same memory and the same kanban board. Install with `uv tool install celeborn`, then `celeborn init`.
Frequently asked
- How is this different from a vector database or RAG memory?
- Celeborn Code's memory is authored markdown you can read and edit, with a full-text index for recall. It is not a black-box embedding store: the Hot tier is a human-curated headline, and search returns exact snippets from files, so you always know what the agent is reading.
- Which agents does it support?
- Claude Code, Codex, and Grok drive the same .context/ memory. The CLI installs hooks for each; the memory format is shared across them.