Anthracode

Why files, not vectors

Vector stores are good at fuzzy recall but bad at one thing humans need: knowing what the agent thinks it knows. A markdown directory is legible, diffable, and version-controllable. If a memory is wrong, you delete the file.

Memory should remain useful when read by a human.

Layout

~/filesystem

.anthracode/memory/
  user/            preferences and identity (one file per topic)
  project/         per-project facts and conventions
  shared/          team-shared, safe to commit
  scratch/         volatile working memory, auto-trimmed

What an entry looks like

~/.anthracode/memory/user/style.md

---
type: user
tags: [style, preferences]
created: 2026-04-12
---

Prefers terse responses. Skips trailing summaries.
Pushes back when implementation drifts beyond the requested change.

Lifecycle

Session start. The agent scans memory dirs and selects entries it judges relevant to the current task.
During the session. New entries are written when something is learned that should outlive this run.
On exit. Volatile scratch/ entries are trimmed; everything else stays.

Editing memory by hand

Memory is just markdown. $EDITOR .anthracode/memory/user/style.md works. The agent re-reads on the next turn — there's no cache to bust.

What NOT to put in memory

Code, structure, conventions — derivable by reading the project.
Git history — git log is authoritative.
Secrets — memory is plain files; use OS keychains or env vars.
One-off task state — that's what tasks are for.