naive-memory

Naive Memory

Overview

Persist useful, repository-specific knowledge discovered during a session into a shared memory file at .genai/memory.md in the repository root. This file serves as a durable scratchpad that any LLM agent can read at the start of future sessions to avoid rediscovering the same information.

When to Invoke

• At the end of a working session, before closing out
• When explicitly asked to "remember" or "save notes"
• After resolving a non-obvious problem where the solution would help future sessions

Workflow

1. Reflect

Ask yourself:

Is there any generalizable knowledge specific to this repository that I wish I'd known prior to this conversation that would have made work more efficient?

Only proceed if the answer is yes.

2. Read existing memory

Read .genai/memory.md from the repository root. If the file does not exist, create it with a # Notes to self heading.

3. Evaluate before writing

Add a new entry under # Notes to self IF AND ONLY IF all of the following are true:

1. The knowledge is not already documented in the file
2. The knowledge is generally useful for future work on this repository
3. The knowledge is unlikely to change soon (avoid transient details)
4. The knowledge is specific to this repository (not general programming knowledge)

4. Write the entry

Each entry must include:

• Timestamp: ISO 8601 date (YYYY-MM-DD)
• Model name: The model that authored the entry (e.g., claude-opus-4-6)
• Content: Concise, accurate, actionable knowledge

Format:

## <Topic>

_Added: YYYY-MM-DD by <model-name>_

<knowledge>

5. Maintain the file

• Keep the file concise — remove or update entries that are outdated or proven wrong
• Deduplicate — if new knowledge supersedes an existing entry, replace rather than append
• Organize semantically by topic, not chronologically

Anti-patterns

• Do NOT document session narratives ("Today I fixed bug X")
• Do NOT include secrets, credentials, or API keys
• Do NOT add general programming knowledge ("Python uses indentation")
• Do NOT add information that is already obvious from the codebase structure
• Do NOT add entries on every session — only when genuinely useful knowledge was discovered