What
A comprehensive persistent memory system that lets Claude remember insights, patterns, and learnings across conversations, stored in MEMORY.md files within configurable directory scopes.
Usage
Agent memory is configured through a new interactive dialog accessible during onboarding or settings. Users choose one of:
- User scope (
~/.claude/agent-memory/) -- general learnings across all projects - Project scope (
.claude/agent-memory/) -- project-specific, shared via version control - Local scope (
.claude/agent-memory-local/) -- project-specific, not checked into version control - None -- no persistent memory
Details
- Memory is stored in
MEMORY.mdfiles that are automatically loaded into the system prompt - Claude is instructed to organize memory semantically by topic, not chronologically
- Records insights about problem constraints, strategies that worked or failed, and lessons learned
- Claude proactively updates memory as it discovers useful patterns during work
- MEMORY.md is truncated at 200 lines with a warning when exceeded, encouraging users to compact or reorganize
- When MEMORY.md is empty, Claude is prompted to start recording learnings for future sessions
- Can be disabled with
CLAUDE_CODE_DISABLE_AUTO_MEMORY=1environment variable
Evidence
Persistent Agent Memory feature with configuration UI (search for "Persistent Agent Memory", "Configure agent memory", "Organize memory semantically by topic", and "no persistent memory"). Gated by tengu_oboe feature flag (default !1 = false).
Strings lifted out of the shipped bundle, so the claim above can be checked against them.