Memory Recall Plugin

Multi-dimension context recall for Claude Code. Automatically surfaces relevant memories, skills, tools, and agent types on every user message.

Features

• 4 hooks: recall (UserPromptSubmit/SubagentStart), memory save (Stop), pair programmer (PostToolUse)
• 4 recall dimensions: memory files, skills, tools (MCP + deferred), agent types
• 3 backends per dimension: reminder (zero-cost), agentic (Haiku selection), embedding (local RAG)
• Pair programmer: evaluates agent actions against user preferences, past experience, strategic direction
• Memory save: auto-saves conversation knowledge to Memory Bank after each turn
• Configurable sync/async: each hook can run synchronously or asynchronously
• 4 skills: /dream (consolidation), /remember (quick save), /setup (config), /diagnose (troubleshooting)

Installation

claude plugin marketplace add t2ance/memory-recall-plugin
claude plugin install memory-recall@memory-recall

Then configure via /setup or manually in ~/.claude/settings.json:

{
  "pluginConfigs": {
    "memory-recall@memory-recall": {
      "options": {
        "memory": "agentic",
        "skills": "agentic",
        "tools": "agentic",
        "agents": "agentic"
      }
    }
  }
}

Each dimension accepts: "off", "reminder", "agentic", or "embedding".

Configuration Reference

Recall options:

Option	Description	Default
memory / skills / tools / agents	Backend per dimension: off, reminder, agentic, embedding	reminder (memory), off (others)
agentic_mode	parallel (one call/dim) or merged (single call)	parallel
{dim}_input	What selector sees: title_desc or full	title_desc
{dim}_output	What gets injected: title_desc or full	full (memory), title_desc (others)
model	Agentic model: haiku / sonnet / opus	haiku
context_messages	Recent messages for search context	5
context_max_chars	Max chars of conversation context	2000
max_content_chars	Global cap on total injected content	9000
recall_effort	Effort for recall calls: low or ""	""

Embedding options:

Option	Description	Default
embedding_model	HuggingFace model name	intfloat/multilingual-e5-small
embedding_python	Python path with sentence-transformers	~/miniconda3/envs/memory-recall/bin/python
embedding_threshold	Cosine similarity threshold	0.85
embedding_top_k	Max results per dimension	3
embedding_device	cpu or cuda	cpu

Memory save options:

Option	Description	Default
auto_save_enabled	Enable auto-save after each turn	true
auto_save_targets	native (project), global, or both	native
auto_save_context_turns	Conversation turns for analysis	3
auto_save_effort	Effort level for save calls	""

Pair programmer options:

Option	Description	Default
pp_enabled	Enable pair programmer	false
pp_model	Model for evaluation	haiku
pp_sample_rate	Probability of evaluating each tool call (0-1)	1.0
pp_cooldown_s	Min seconds between evaluations	0
pp_context_messages	Recent messages for trajectory	5
pp_context_max_chars	Max conversation context chars	3000
pp_effort	Effort level	""
pp_max_tool_input_chars	Max tool input chars in trajectory	2000
pp_max_tool_output_chars	Max tool output chars in trajectory	1000
pp_max_recall_files	Max memory files to recall	5
pp_max_memory_file_chars	Max chars per recalled memory file	2000

Async options:

Option	Description	Default
recall_async	Run recall hook asynchronously	false
memory_save_async	Run memory save hook asynchronously	true
pp_async	Run pair programmer hook asynchronously	true

How It Works

Recall (UserPromptSubmit / SubagentStart)

On every user message and sub-agent spawn, the hook:

1. Discovers available resources per enabled dimension (file scan + hardcoded fallback)
2. Recalls relevant items using the configured backend (parallel for agentic)
3. Injects results as additionalContext into the model's context

Sub-agents and teammates also receive recall context. On SubagentStart, the hook extracts the parent agent's prompt from the transcript and runs the full recall pipeline.

Memory Save (Stop)

After each assistant turn, the hook:

1. Extracts recent conversation turns from transcript
2. Calls Haiku to decide what knowledge to persist (ADD/UPDATE/DELETE/NOOP)
3. Writes memory files and updates MEMORY.md index