---
name: save
description: "Extract knowledge from the current conversation and write it to gestalt."
disable-model-invocation: true
---
# Save
Extract knowledge from the current conversation and write it to gestalt. Also triggered automatically when context is getting full.
## When to Run
- User invokes `/save` explicitly (manual override)
- The conversation is long and context is approaching its limit — save proactively before the chat gets summarized so nothing is lost
- **Note:** Knowledge is also captured automatically by the Stop hook at session end. Manual `/save` is only needed for mid-session captures.
## Process
### Step 1: Extract Everything Worth Keeping
Scan the full conversation for:
- Corrections the user made (mistakes to avoid)
- Conventions or patterns that were established
- How-tos that were figured out (setup, run, test, deploy)
- Architectural decisions or design rationale
- Debugging insights (what worked, what didn't)
- Data flow discoveries (what connects to what)
- Cross-repo relationships uncovered
- External source references (Notion pages, Linear issues, Slack threads cited)
- Anything the user explicitly asked to remember
### Step 2: Check Existing Coverage
Follow the gestalt grounding procedure in `gestalt/.cursor/references/gestalt-grounding.md` (Quick Lookup). For each piece of knowledge extracted, check if an existing entry covers it. If so, you will update that entry. If not, create a new one.
### Step 3: Write or Rewrite
Follow the knowledge write conventions in `gestalt/.cursor/references/knowledge-write.md` — use the frontmatter template, linking rules, and write philosophy defined there.
**For a rule** — create or update `gestalt/rules/<slug>.mdc` with `description`, `globs`, `alwaysApply: false`.
**For knowledge** — create or update `gestalt/knowledge/<slug>.md`.
### Step 4: Register External Sources
If the conversation referenced external sources (Notion pages, Linear issues, Slack discussions), record them in the entry's `## Data Flow` or `## Relationships` section and update `gestalt/SOURCES.md`.
### Step 5: Validate
Check that all `wikilinks` in new/updated entries reference existing entries (check MANIFEST.md). If a link target doesn't exist, either create the entry or remove the link.
### Step 6: Regenerate Indices
Execute the full Post-Write Sequence per `gestalt/.cursor/references/knowledge-write.md` §Post-Write Sequence (indices + semantic index + Graphiti). All parts are mandatory.
### Step 6.5: Rebuild Semantic Index (if available)
(Covered by Post-Write Sequence §b above.)
### Step 6.6: Feed Temporal Graph (if available)
Run `tools/gestalt-graphiti-sync.sh <slug>` for each entry written this session (covered by Post-Write Sequence §c above — this is the same hash-gated call, not a separate `add_memory`).
### Step 7: Confirm
Tell the user what was saved/updated. If this was an auto-save before context limit, note: "Saved conversation knowledge to gestalt before context limit."
**Note:** The Stop hook (gestalt-stop.sh) also captures session knowledge at session end; it deduplicates against MANIFEST.md before writing, so a manual /save mid-session will not produce duplicates.
## Verification Gate
Execute every step above — none are optional.
Before finishing: confirm all steps completed and outputs match the required formats.
## Evals — see EVALS.md (dev-time only)
