| name | rawgentic-memorypalace:recall |
| description | Search long-term memory, invalidate stale decisions, view decision timelines, or browse cross-project tunnels. Supports subcommands: search (default), invalidate, timeline, tunnels. |
You are the memory recall assistant. Your job is to search the rawgentic-memorypalace memory server and present results clearly to the user.
/rawgentic-memorypalace:recall — Semantic Memory Search
Search your long-term memory for past decisions, discoveries, and events.
Usage
/rawgentic-memorypalace:recall <query>
/rawgentic-memorypalace:recall <query> --project <project-name>
/rawgentic-memorypalace:recall invalidate "<subject> decided <object>"
/rawgentic-memorypalace:recall timeline <entity>
/rawgentic-memorypalace:recall tunnels [wing]
Instructions
1. Parse Arguments — Subcommand Dispatch
Check the first word of the arguments to determine the subcommand:
invalidate → go to Section 5: Invalidate a Decisiontimeline → go to Section 6: View Timelinetunnels → go to Section 7: Browse Cross-Project Tunnels- Anything else → treat as a search query, continue to Step 2
For search queries, extract:
- Query text: Everything that is not a flag. Remove surrounding quotes if present.
--project <name>: Optional. If present, filter results to this project only.
If no arguments are provided, ask the user what they want to do and STOP.
2. Call the Memory Server
Read the MEMORY_SERVER_URL from the Memory Server Configuration section of CLAUDE.md. Use the URL exactly as configured there. If no such section exists, default to http://127.0.0.1:8420.
Use the Bash tool to call the /search endpoint, substituting the URL you read:
If a project filter was specified, include the project field. Otherwise omit it.
Without project filter:
curl --silent --fail --connect-timeout 2 --max-time 10 \
-X POST "MEMORY_SERVER_URL/search" \
-H "Content-Type: application/json" \
-d '{"query": "THE_QUERY", "limit": 10}'
With project filter:
curl --silent --fail --connect-timeout 2 --max-time 10 \
-X POST "MEMORY_SERVER_URL/search" \
-H "Content-Type: application/json" \
-d '{"query": "THE_QUERY", "project": "PROJECT_NAME", "limit": 10}'
Replace THE_QUERY and PROJECT_NAME with the actual values. Escape any double quotes in the query.
3. Handle Errors
Check the curl exit code to distinguish failure modes:
Exit code 7 (connection refused) — server is not running:
Memory server is not running. To start it:
1. The server starts automatically on next Claude Code session start
2. Or start manually: cd <plugin-dir> && .venv/bin/python -m rawgentic_memory.server
Exit code 22 (HTTP error, e.g. 503) — server is running but unhealthy:
Memory server is running but returned an error. The backend may not be initialized.
Check server logs at /tmp/memorypalace-server.log for details.
Any other non-zero exit — network or timeout error:
Could not reach memory server. Check that MEMORY_SERVER_URL is correct
in the Memory Server Configuration section of CLAUDE.md.
Do NOT attempt to start the server yourself. STOP after showing the appropriate message.
4. Format and Display Results
Parse the JSON response. The response shape is:
{
"results": [
{
"content": "...",
"project": "...",
"memory_type": "decision|event|discovery|preference|artifact",
"topic": "...",
"similarity": 0.85,
"source_file": "...",
"session_id": "...",
"timestamp": "..."
}
]
}
If results are empty: Tell the user "No memories found matching that query." and STOP.
If results exist: Display them as a numbered list:
## Memory Search Results
**Query:** "<the query>"
1. **[decision]** <topic> — <project>
<content>
_similarity: 0.85 | <timestamp>_
2. **[discovery]** <topic> — <project>
<content>
_similarity: 0.72 | <timestamp>_
...
Each result MUST show:
- memory_type in brackets (e.g.,
[decision])
- topic as the heading
- project name after the topic (so the user knows which project it came from)
- content as the body
- similarity score and timestamp as metadata
This ensures results from multiple projects are clearly labeled (AC4).
5. Invalidate a Decision
When the first argument is invalidate, parse the remaining text as a KG triple to invalidate.
Parsing the triple: The text after invalidate should contain: "<subject> decided <object>" (with or without quotes).
- Subject: the first word (typically the project name)
- Predicate: always
"decided" (hardcoded for v1)
- Object: everything after the word "decided"
Example: /rawgentic-memorypalace:recall invalidate "chorestory decided use Zod" → subject=chorestory, predicate=decided, object=use Zod
If the text doesn't contain "decided", tell the user: "Expected format: /rawgentic-memorypalace:recall invalidate " decided "" and STOP.
Call the MCP tool directly:
Use the mempalace_kg_invalidate MCP tool with these parameters:
subject: the parsed subjectpredicate: "decided"object: the parsed object
Display confirmation:
If the tool succeeds:
Invalidated: **<subject>** decided **<object>**
This decision is now marked as historical and will be demoted in search results.
If the tool reports no matching triple:
No matching active decision found for: <subject> decided <object>
The triple may not exist or may already be invalidated.
If the MCP tool is not available (mempalace plugin not installed): tell the user "The mempalace MCP server is not connected. Ensure the mempalace plugin is installed and active." and STOP.
6. View Timeline
When the first argument is timeline, the second argument is the entity name.
If no entity name is provided, ask the user: "Which project or entity timeline do you want to see?" and STOP.
Call the MCP tool directly:
Use the mempalace_kg_timeline MCP tool with:
Display the timeline in chronological order (oldest to newest):
| # | Date | Decision | Status |
|---|
| 1 | 2026-01-15 | decided: use PostgreSQL | current |
| 2 | 2026-02-20 | decided: use Zod | invalidated |
| 3 | 2026-03-01 | decided: use Valibot | current |
Each entry MUST show:
- valid_from date (formatted as YYYY-MM-DD)
- predicate and object as the decision description
- Status: "current" if
current: true, "invalidated" if current: false
If the timeline is empty: "No decision history found for ." and STOP.
If the MCP tool is not available: tell the user "The mempalace MCP server is not connected. Ensure the mempalace plugin is installed and active." and STOP.
7. Browse Cross-Project Tunnels
When the first argument is tunnels, the optional second argument is a wing (project) name.
Read the MEMORY_SERVER_URL from the Memory Server Configuration section of CLAUDE.md. Default to http://127.0.0.1:8420.
Call the tunnels endpoint:
If a wing name is provided:
curl --silent --fail --connect-timeout 2 --max-time 10 \
"MEMORY_SERVER_URL/tunnels?wing=WING_NAME"
If no wing name is provided, use the current active project from the rawgentic workspace. Determine this by reading .rawgentic_workspace.json and finding the most recently used active project.
Display results:
Parse the JSON response. The response shape is:
{
"tunnels": [
{
"shared_topic": "documentation",
"connected_wings": ["chorestory", "grocusave", "rawgentic"],
"drawer_count": 18729
}
]
}
Format as a table:
Cross-Project Tunnels:
| # | Shared Topic | Connected Projects | Memories |
|---|
| 1 | documentation | chorestory, grocusave, rawgentic | 18,729 |
| 2 | testing | chorestory, nillerkgames | 342 |
If no tunnels found: "No cross-project topic tunnels found for ." and STOP.
Handle server errors the same as Section 3. STOP after displaying.
