Cogmemai Memory

CogmemAi Memory Management

CogmemAi gives your Ai assistant persistent memory across sessions. This skill teaches you how to use it effectively — when to save, what to save, and how to keep your memory organized.

Core Workflow

Session Start

Always call get_project_context before responding to the user's first message. This loads your most important memories from previous sessions. Pass an optional context parameter describing the current task for better relevance.

While Working

Save memories when you learn something important:

• Architecture decisions and tech stack details (importance: 8-10)
• User preferences for coding style, tools, workflow (importance: 7-9)
• Bug fixes, gotchas, and workarounds (importance: 6-8)
• Key file paths and project structure (importance: 7-9)
• Patterns and conventions used in the codebase (importance: 6-8)
• Dependency versions and constraints (importance: 5-7)

After Significant Exchanges

Use extract_memories after important conversations instead of manually saving each fact. It automatically identifies what's worth remembering:

• After debugging a tricky bug — captures the root cause and fix
• After discussing architecture — captures decisions and trade-offs
• After the user explains their preferences — captures all stated preferences
• After a planning session — captures key decisions and constraints

This is faster and more thorough than manual save_memory calls for multi-fact exchanges.

Session End

When the session ends or the user says goodbye, save a session summary capturing what was accomplished, decisions made, and next steps.

When to Save a Memory

Save immediately when:

• The user tells you a preference ("always use tabs", "never auto-commit")
• You discover an architecture decision ("auth uses JWT", "DB is PostgreSQL 15")
• You fix a bug and the fix is non-obvious
• You learn a project convention ("components use PascalCase", "tests in tests/")
• The user makes a decision ("we'll use Redis for caching")

Do NOT save:

• Temporary task details (what you're working on right now)
• Information already in CLAUDE.md or project docs
• Speculative conclusions from reading a single file
• Duplicate information you've already saved

Importance Scoring Guide

Memory Types

Choose the right type for better retrieval:

• identity — Who the user is, their role, team
• preference — Coding style, tool choices, conventions
• architecture — System design, tech stack, file structure
• decision — Why X was chosen over Y
• bug — Known issues, fixes, workarounds
• dependency — Version constraints, package notes
• pattern — Reusable patterns, conventions
• context — General project context
• session_summary — What was accomplished in a session
• task — Persistent tasks with status and priority tracking
• correction — Wrong approach to right approach patterns
• reminder — Next-session nudges that auto-expire

Scoping

• project (default) — Specific to this codebase. Architecture, bugs, file paths.
• global — Applies everywhere. User preferences, identity, tool choices.

Use promote_memory to move a project memory to global scope when you discover it applies universally.

Workflows

Project Onboarding

When starting work on a new project for the first time:

1. Call get_project_context to check for existing memories
2. If none exist, read the project's README, package.json, or equivalent
3. Use ingest_document to extract key facts from documentation
4. Save 3-5 high-importance memories about the tech stack and structure
5. Save the user's stated preferences for this project

Memory Health Check

Periodically (or when the user asks), review memory quality:

1. Check the health_score from get_project_context — a 0-100 score showing overall memory health
2. Call get_analytics to see usage patterns — it automatically tunes importance and archives stale memories
3. Review "never recalled" memories — delete ones that aren't useful
4. Use consolidate_memories to merge related facts into comprehensive summaries
5. Use link_memories to connect related facts
6. Use get_stale_memories to find outdated memories that need review

Task Tracking

Use tasks for work that spans multiple sessions:

1. save_task to create a task with status (todo, in_progress, done, blocked) and priority
2. get_tasks at session start to see what's pending
3. update_task as you make progress
4. Tasks persist across sessions and compaction — nothing gets lost

Corrections and Reminders

• Use save_correction when you discover a wrong approach — stores the wrong-to-right pattern so the mistake isn't repeated
• Use set_reminder to surface a nudge at the start of the next session (auto-expires after delivery)

Context Recovery After Compaction

When Claude Code compacts your context (you'll see a summary of the previous conversation):

1. Call get_project_context immediately to reload memories
2. Pass the current task as context for better relevance
3. Continue working — your memories bridge the gap

Searching for Specific Knowledge

When you need to find something specific from past sessions:

1. Use recall_memories with a natural language query
2. Filter by memory_type or category if you know the type
3. Results are intelligently ranked for maximum relevance

Best Practices

• Keep memories concise — Complete sentences, 1-2 lines each. Not paragraphs.
• Use descriptive subjects — "auth_system", "database_setup", "css_conventions" — not "note1"
• Save immediately — Don't wait until the end of a session. Save as you learn.
• Use extract_memories for multi-fact exchanges — When a conversation covers multiple important topics, use extract_memories instead of multiple save_memory calls. It's faster and catches facts you might miss.
• Update, don't duplicate — Use update_memory to revise existing memories rather than creating duplicates. The system detects conflicts automatically.
• Use bulk_update for maintenance — When cleaning up memory types, categories, or importance scores across many memories, use bulk_update instead of individual calls.
• Use tags for grouping related memories across types (e.g., tag "auth" on architecture, bug, and decision memories about authentication)

Common Mistakes

• Setting everything to importance 10 — Reserve 9-10 for core architecture only
• Saving raw code as memory content — Save the fact, not the code ("Login component uses useEffect for auth check", not the actual useEffect code)
• Forgetting to call get_project_context at session start — This is the single most important action for continuity
• Saving session-specific state — "Currently working on the login bug" will be stale next session