Organizational Memory
Codumentor can build and maintain a shared knowledge base from conversations across your team. Insights, conventions, decisions, and patterns are automatically captured, deduplicated, and surfaced when relevant.
This feature may not be available in all setups. If you don't see memory cards in your conversations, it may not be enabled by your administrator.
How it works
- Automatic extraction — After each conversation, a background process analyzes the exchange and extracts useful insights (naming conventions, architectural decisions, workarounds, etc.)
- Deduplication — The system checks if similar knowledge already exists and updates or merges rather than creating duplicates
- Retrieval — Before the assistant answers your question, it searches the knowledge base for relevant entries and uses them as context
- Version control — All knowledge is stored as Markdown files in a dedicated git repository, with full history and diffs
Rapid Direction Cues
When the assistant finds relevant knowledge, it receives these as Rapid Direction Cues — brief, contextual hints injected before it responds. You may see references to these in the assistant's answers, such as "Based on existing team knowledge..." or similar phrases.
This means the assistant's answers are grounded in what your team actually knows and has decided, not just generic patterns.
Voting and feedback
When the assistant surfaces a memory, you'll see voting controls:
| Action | Effect |
|---|---|
| Like | Reinforces the memory — it will be prioritized in future retrieval |
| Dislike | Signals the memory may be wrong or unhelpful — provide a reason if possible |
| Flag | Escalates the memory for human review |
Your votes directly improve the knowledge base for everyone on the team.
Memory cards
Memory-related activity appears as cards in the conversation transcript:
- Memory stored — a new insight was extracted and saved
- Memory retrieved — existing knowledge was found and used to inform the response
These cards show summaries of what was stored or retrieved.
Confidence and seniority
The system assigns confidence scores to memories based on the author's role. More senior contributors' insights are weighted higher in search results. This happens automatically — you don't need to configure anything.
What gets stored
The system extracts:
- Naming conventions and coding standards
- Architectural decisions and their rationale
- Undocumented workarounds and gotchas
- Cross-repository dependencies and relationships
- Historical context about why things were built a certain way
It does not store sensitive information like passwords, API keys, or personal data (though this depends on your organization's configuration).