shane9coy Remember — Store Memory
When This Fires
- User types
/remember - User says "remember this", "save that", "note that down"
- Auto mode: agent detects a natural breakpoint (topic switch, milestone, wrap-up)
What To Do
Step 1: Determine what type of memory this is
Read the conversation and classify:
| If the conversation contains... | Write to... |
|---|---|
| User preference about how the agent should behave ("be more direct", "stop asking if I want to proceed") | ~/.katana/memory/soul.md → Learned Behaviors section |
| What user is currently working on, struggling with, goals, life events, mood, accomplishments | ~/.katana/memory/user.md → appropriate section (Current Focus, Recent Context, Wins, etc.) |
| Work done on a project, code changes, architecture decisions, tasks completed | ~/.katana/memory/work.md → new entry at top |
If the session contains multiple types, write to multiple files.
Step 2: Summarize
- Strip all raw code blocks longer than 5 lines
- Strip tool call outputs and error logs
- Strip repetitive back-and-forth troubleshooting
- Keep: decisions made, problems solved, features built, user preferences stated
- Condense to 1-3 sentences for work.md entries
- For user.md and soul.md: extract the specific context, pattern, or preference (not a session summary)
Step 3: Detect project context
For work.md entries:
- Check if user specified a project:
/remember --project myapp - Check current working directory for git repo name: run
basename $(git rev-parse --show-toplevel 2>/dev/null)or check the current directory name - If neither available, ask the user: "What project was this for?"
- Fallback: tag as
general
Step 4: Write
For work.md — prepend new entry at the top of the file (after the header comments):
## YYYY-MM-DD — project-tag
Summary sentence(s) here.
For user.md — find the right section header and append the new context:
## Recent Context
- YYYY-MM-DD: Description of what happened or what the agent learned.
For soul.md — append to Learned Behaviors:
## Learned Behaviors
- Existing behavior
- NEW: User wants responses under 5 sentences unless asked for detail
Step 5: Confirm (manual mode only)
Check ~/.katana/katana.toml for the memory mode setting.
If memory.mode = "manual" (default), show the user what will be saved and where:
📝 Memory save:
→ work.md: "Designed v1 memory system with three files and two commands."
Tagged: katana-agent | 2026-02-22
Save this? (y/n/edit)
If memory.mode = "auto", skip confirmation and write directly.
Step 6: Acknowledge
After saving, briefly confirm:
✓ Saved to work.md (katana-agent, 2026-02-22)
Auto Mode Rules
When memory.mode = "auto" in katana.toml:
Save to work.md when:
- The user switches to a different project or topic
- Something major gets completed (tests pass, feature works, bug fixed)
- 20+ exchanges have happened without a save
- The user says goodbye, thanks, or wraps up
Save to user.md when:
- The user shares how they're feeling, what they're struggling with, or goals
- The user mentions a life event, big win, or setback
Save to soul.md when:
- The user corrects the agent's behavior or states a preference
Retention
work.md has a max entry limit (default 200, configurable in katana.toml). When work.md exceeds this count, trim the oldest entries (bottom of file) on the next write. soul.md and user.md don't need retention — they're living documents that get updated in place, not appended infinitely.