Rememora

Rememora — Claude Code Plugin

Persistent cross-agent memory for Claude Code. Automatically saves decisions, bug fixes, patterns, and entity knowledge across sessions. Semantically retrieves relevant context when you need it.

What it does

• SessionStart hook: Loads project context from rememora automatically
• Stop hook: Curates memories from the transcript after each agent turn (one in-flight curate per session, via a kernel-level concurrency gate)
• SessionEnd hook: Closes the active rememora session and runs a final curation pass
• Model-invoked save skill: Claude autonomously saves knowledge when it makes decisions, fixes bugs, or discovers patterns
• Model-invoked search skill: Claude autonomously searches memory before implementing or when encountering unfamiliar code
• /rememora command: Manually trigger memory save or search

Install

# 1. Add the Rememora marketplace
claude plugin marketplace add Rememora/rememora

# 2. Install the plugin
claude plugin install rememora@rememora

# For project-wide install (shared via git):
claude plugin install rememora@rememora --scope project

Requirements

• rememora CLI installed and on PATH (cargo install rememora or via Homebrew)
• A registered project: rememora project add <name> --path <cwd>

Escape hatches

If a Rememora hook misbehaves and you need to turn everything off quickly, set the kill-switch in the shell that launches Claude Code:

export REMEMORA_DISABLE_HOOKS=1

All four hook scripts (session-start.sh, session-end.sh, stop-curate.sh, prompt-search.sh) check this env var at the top and early-exit 0 when set. unset REMEMORA_DISABLE_HOOKS to re-enable.

Other tunables:

• REMEMORA_CURATE_COOLDOWN_SECS (default 300): minimum seconds between curate runs per session. Set to 0 to disable the frequency gate. The kernel-level pgrep concurrency gate is independent and always active.
• REMEMORA_CURATE_CHILD: set internally by rememora curate on its claude -p subprocesses so the entire hook chain is a no-op inside curator children (no spurious session rows, no context injection, no recursive curate). Not intended for user override.

How it works

1. On session start, the hook runs rememora context --auto and injects prior knowledge
2. During work, Claude autonomously saves when it detects:
   • Architectural or design decisions
   • Non-trivial bug fixes
   • Codebase patterns or conventions
   • Important entities (services, APIs, configs)
3. Before implementation, Claude autonomously searches for relevant prior knowledge
4. After each agent turn, the Stop hook forks rememora curate against the session transcript to extract anything Claude missed. The curate process is fully detached — launched in its own session via setsid (or nohup + disown on stock macOS) with stdin/stdout/stderr redirected to /dev/null, so it cannot hold the hook's pipe and block Claude Code waiting for EOF. At most one curate runs in-flight per session (enforced by a pgrep-based concurrency gate); a secondary REMEMORA_CURATE_COOLDOWN_SECS (default 300) frequency gate rate-limits consecutive runs
5. On session end, the hook closes the active rememora session and runs a final curation pass so the tail of the session is never lost

Plugin structure

plugin/
├── hooks/
│   └── hooks.json              # SessionStart + Stop + SessionEnd hooks
├── scripts/
│   ├── session-start.sh        # Load context + start session
│   ├── stop-curate.sh          # Fork `rememora curate` per agent turn
│   └── session-end.sh          # End active session + final curation pass
├── skills/
│   ├── rememora-save/
│   │   └── SKILL.md            # Model-invoked: autonomous save
│   ├── rememora-search/
│   │   └── SKILL.md            # Model-invoked: autonomous search
│   └── rememora-init/
│       └── SKILL.md            # User-invoked: /rememora
└── README.md