obra/remembering-conversations

Remembering Conversations

Core principle: Search before reinventing. Searching costs nothing; reinventing or repeating mistakes costs everything.

Mandatory: Search Historical Memory

YOU MUST search historical memory for any historical search.

Announce: "Searching past conversations for [topic]."

Claude Code

Use the Task tool with subagent_type: "search-conversations":

Task tool:
  description: "Search past conversations for [topic]"
  prompt: "Search for [specific query or topic]. Focus on [what you're looking for - e.g., decisions, patterns, gotchas, code examples]."
  subagent_type: "search-conversations"

Codex

If a search-conversations agent is available, dispatch it with the same prompt. If not, use the MCP tools directly:

1. Search with the episodic-memory search tool
2. Read the top 2-5 results with the episodic-memory read tool
3. Synthesize findings in your response
4. Include source pointers so the user can inspect the original conversations

The search workflow will:

1. Search with the search tool
2. Read top 2-5 results with the read tool
3. Synthesize findings (200-1000 words)
4. Return actionable insights + sources

Saves 50-100x context vs. loading raw conversations.

When to Use

Use this whenever the current task would benefit from information you may have learned before, even if the user did not explicitly ask you to search.

When past experience may help:

• You need to recall decisions, rationale, patterns, solutions, pitfalls, or project context from earlier work
• A task resembles something you've solved, debugged, reviewed, released, or planned before
• You need to repeat a workflow or process that may have prior gotchas or established steps

When you're stuck:

• You've investigated a problem and can't find the solution
• Facing a complex problem without obvious solution in current code
• Need to follow an unfamiliar workflow or process

When historical signals are present:

• User says "last time", "before", "we discussed", "you implemented"
• User asks "why did we...", "what was the reason..."
• User says "do you remember...", "what do we know about..."

Before answering from uncertainty:

• Before guessing from memory or saying "I don't know" about something that may have been learned in a past conversation, search memory unless the current conversation already answers it

Don't search first:

• For current codebase structure (use Grep/Read to explore first)
• For info in current conversation
• Before understanding what you're being asked to do

Direct MCP Tool Access

Use these directly when a search agent is unavailable or the current harness does not support agent dispatch:

• mcp__plugin_episodic-memory_episodic-memory__search
• mcp__plugin_episodic-memory_episodic-memory__read

When using MCP tools directly, keep context small: search first, then read only the top 2-5 relevant conversations or line ranges.

See MCP-TOOLS.md for complete API reference if needed for advanced usage.