memory-rlm

MemoryRLM

Persistent project memory for Claude Code. A Rust MCP server that transparently indexes your conversation context and code changes, then re-injects the most relevant information when context is lost to compaction.

Why this exists

Claude Code sessions have finite context windows. As conversations grow through requests, code edits, debugging, and planning, older context gets compressed or discarded. You lose decisions, rationale, and the thread of what you were doing.

ClaudeRLM fixes this by indexing everything as it happens and surgically re-injecting what matters when context is lost.

How it works

The core idea is borrowed from Recursive LLMs (Zhang et al., 2025), which showed that LLMs can process arbitrarily long inputs by treating them as external data to be programmatically queried rather than consumed all at once. Our insight: conversation context grows incrementally, so it can be indexed at write-time as each turn happens, rather than processed at read-time when it's already too late. This makes indexing nearly free -- each hook call takes milliseconds -- and retrieval is a fast SQLite query.

In practice, Claude Code hooks fire on every user prompt, code edit, file read, and bash command. ClaudeRLM captures each event into a local SQLite database with FTS5 full-text search. When compaction is about to happen, a PreCompact hook ensures everything is indexed and creates a checkpoint summary. After compaction, a SessionStart hook queries the index and injects the most relevant context back into the conversation -- ranked by recency, type importance, and file affinity -- so Claude picks up right where it left off.

Features

• Passive indexing -- hooks fire automatically, Claude never needs to decide to use it
• Full-text search over conversation history (SQLite FTS5 with BM25 ranking)
• Code structure indexing via tree-sitter (Rust, Python, TypeScript, JavaScript, Go, C, C++)
• Background file watcher for incremental re-indexing on file changes
• Ranked context injection after compaction (type weight x recency x file affinity)
• Knowledge distillation at session end -- extracts decisions, preferences, conventions, and bug fixes
• LLM-enhanced distillation with Haiku (~1 cent/session) or any OpenAI-compatible endpoint (Ollama for free)
• 4 MCP tools for explicit search when needed: memory_search, memory_symbols, memory_decisions, memory_files
• Cross-session memory -- knowledge persists and is injected at the start of every new session

Install

Plugin install (recommended)

Install as a Claude Code plugin for one-command setup with automatic binary management:

/plugin marketplace add dullfig/claude-plugins
/plugin install memory-rlm

The binary is downloaded automatically on first session start. No PATH configuration needed.

To update, delete the cached binary and restart:

rm -f <plugin-dir>/bin/memory-rlm    # Unix
del <plugin-dir>\bin\memory-rlm.exe  # Windows

Manual install

If you prefer not to use the plugin system, standalone installers are still available:

Linux / macOS:

curl -fsSL https://raw.githubusercontent.com/dullfig/memory-rlm/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/dullfig/memory-rlm/main/install.ps1 | iex

From source (any platform with Rust):

cargo install --git https://github.com/dullfig/memory-rlm.git

If you install manually, you'll need to configure hooks yourself (see Manual hook setup below). Use absolute paths to the binary — Claude Code spawns hook subprocesses without inheriting your shell PATH.

Migrating from manual install to plugin

1. Install the plugin: /plugin marketplace add dullfig/claude-plugins then /plugin install memory-rlm
2. Remove old hooks from ~/.claude/settings.json (any entries referencing memory-rlm)
3. Optionally remove the old binary from ~/.local/bin/memory-rlm or %LOCALAPPDATA%\Programs\memory-rlm\

Your existing database (.claude/memory-rlm.db) is preserved — all your indexed memory carries over.

Manual hook setup

If you installed manually, copy hooks.example.json into your project's .claude/settings.json (or merge into your existing settings).

Important: Use the absolute path to memory-rlm in hook commands. Replace /path/to/memory-rlm with the actual install location (e.g., ~/.local/bin/memory-rlm on Linux/macOS, C:/Users/YOU/AppData/Local/Programs/memory-rlm/memory-rlm.exe on Windows):