memory

Purpose

Persist project knowledge across sessions. Record decisions, capture learnings, store code patterns, and retrieve context when needed. Prevents knowledge loss and repeated mistakes.

When to Invoke

• •Recording decisions, learnings, gotchas, or artifacts
• •User says "remember", "store", "save", "record this"
• •User asks "what did we decide about...", "why did we..."
• •Before starting significant work (check gotchas first!)
• •After fixing non-obvious bugs or making architectural decisions
• •When deliberating on complex choices (use think commands)
• •User says "brainstorm", "weigh options", "think through", "pros and cons"
• •When you need documented chain-of-thought before committing to a decision

Usage

The memory skill is a TypeScript CLI that runs on Bun.

Prerequisites

Install Bun (if not already installed): Visit https://bun.sh/docs/installation or run:

curl -fsSL https://bun.sh/install | bash

Setup

From the plugin installation directory, run:

bun link

This creates a global memory command. The plugin directory is typically: ~/.claude/plugins/cache/local-memory-plugin/claude-memory-plugin/1.0.0/

Invocation

memory <command> [options]
memory help                    # Quick reference
memory help --full             # Detailed documentation
memory <command> -h            # Command-specific help

Command Reference

Run memory help for quick reference, or memory <command> -h for command-specific help.

Think Command Flags (v1.1.0)

Examples:

# AI-generated thought with auto-selected style
memory think add "Review security implications" --call claude --auto

# Cross-provider calling
memory think add "Analyse performance" --call codex --model gpt-5-codex
memory think counter "Consider alternatives" --call gemini

# With explicit style and agent
memory think add "Deep analysis" --call claude --style Devils-Advocate --agent security-reviewer

Memory Injection (v1.1.0)

The PostToolUse hook can inject relevant memories into agent context. Configure in .claude/memory.local.md:

injection:
  types:
    gotcha:
      enabled: true      # Default: true
      threshold: 0.2
      limit: 5
    decision:
      enabled: false     # Set true to inject decisions
      threshold: 0.35
      limit: 3
    learning:
      enabled: false     # Set true to inject learnings
      threshold: 0.4
      limit: 2
  hook_multipliers:
    Read: 1.0
    Edit: 0.8            # Lower threshold = more sensitive
    Write: 0.8
    Bash: 1.2            # Higher threshold = less noise

Configuration

The plugin works out-of-the-box with sensible defaults. Power users can customise behaviour via .claude/memory.local.md.

Quick Start:

1. •Copy .claude/memory.example.md to .claude/memory.local.md
2. •Edit settings as needed
3. •Add .claude/*.local.md to .gitignore

Available Settings:

Example Configuration:

---
ollama_host: http://192.168.1.100:11434
chat_model: llama3.2
embedding_model: nomic-embed-text
---

See .claude/memory.example.md for full documentation.

Requirements

• •

  Bun - TypeScript runtime (install via bun link from plugin root)

• •

  Ollama (optional) - Required for semantic features:

  • •semantic search by meaning
  • •write --auto-link automatic linking
  • •suggest-links relationship suggestions
  • •refresh --embeddings embedding generation

  Install: ollama pull embeddinggemma (see Ollama docs)

  Without Ollama, semantic features gracefully fall back to keyword search. The plugin continues to function for core operations (memory storage, graph management, keyword search).

Full Documentation

See README.md for detailed documentation on:

• •Architecture and code structure
• •Data contracts and storage layout
• •Quality assessment system
• •Semantic search and auto-linking
• •Thinking sessions workflow
• •Troubleshooting guide