| name | clawql-vault-memory |
| description | Deep ClawQL Obsidian vault workflows using memory_ingest and memory_recall MCP tools—structured insights, wikilink graphs, session threading, verbatim captures, and recall queries. Use when persisting important context, building a knowledge graph, starting tasks that may depend on prior vault notes, or when the user mentions Obsidian, the vault, memory, or long-running project context. |
ClawQL vault memory (deep ingest + recall)
When to apply
- Ingest: After meaningful outcomes—decisions, debugging conclusions, API/env contracts, runbooks, user preferences, cross-feature relationships.
- Recall: Before deep work, when the user references past work, or when vault context could reduce wrong assumptions.
If ClawQL MCP is not configured or CLAWQL_OBSIDIAN_VAULT_PATH is unset, say so briefly and continue without blocking.
Cursor Agent: call_mcp_tool server parameter
Your mcp.json key can stay clawql (that is what Settings / docs refer to). The agent call_mcp_tool API does not accept that bare string: Cursor registers the server under a scoped id derived from where the config lives:
| Config file | Typical call_mcp_tool server value |
|---|
~/.cursor/mcp.json | user-clawql |
${workspaceFolder}/.cursor/mcp.json | workspace-clawql |
There is no supported mcp.json field to rename this to bare clawql for call_mcp_tool—the prefix is host behavior to avoid collisions (e.g. the same key in user vs project config). If both files define clawql, two scoped servers can exist; use the one that matches where you configured ClawQL for this workspace.
Never store secrets: tokens, API keys, passwords, private keys, or raw session cookies. Summarize redacted config instead.
Tool capabilities (what the server actually does)
memory_ingest
| Field | Use |
|---|
title | Required. Drives the H1 and the filename slug under Memory/<slug>.md. Prefer stable, descriptive titles (e.g. gRPC MCP ListTools metadata) so updates append to the same page. |
insights | Primary prose. Markdown is fine. This is where semantic tagging and structure live (see below). |
conversation | Longer transcript or chat summary; stored in a fenced block under Conversation. |
toolOutputs | Verbatim logs: command output, errors, JSON snippets, diffs—each string becomes a block; multiple strings are separated for readability. |
toolOutputsFile | Server-side only: a path (absolute or relative to the ClawQL process cwd) the MCP server reads as UTF-8 and uses as toolOutputs. Use for large bodies so the model only passes a short path in the tool call (not the file contents). Must fall under an allowlist (CLAWQL_MEMORY_INGEST_FILE_ROOTS, or default realpath of cwd). If both toolOutputsFile and toolOutputs are set, the file wins. Disable reads with CLAWQL_MEMORY_INGEST_FILE=0. |
wikilinks | List of other vault note titles (plain names; [[brackets]] optional). Rendered as ## Related with - [[Note Name]] bullets—this is the graph for Obsidian. |
sessionId | Optional label for the section header (threads multi-step work). |
append | Default true: new section appended to an existing file with the same slug. Set false only to replace the file body (rare). |
Dedup: Identical payload hash → ingest is skipped (no duplicate sections).
Frontmatter (automatic): New files get OKF-compatible YAML with required type (default context), title, timestamp, tags (includes clawql-ingest), clawql_ingest, clawql_okf, plus optional ClawQL extensions when passed (correlation_id, worm_ref, agent_id, verdict). The MCP tool accepts type, description, resource, tags, correlationId, wormRef, agentId, verdict — see docs/memory/okf.md. Put extra tagging inside insights as well when useful for humans.
memory_recall
| Field | Use |
|---|
query | Keywords / phrase; tokenized for scoring across vault Markdown. |
limit | Max notes (default from env, often ~10). |
maxDepth | Wikilink hops from keyword hits (default from env). Higher = more graph context. |
minScore | Raise to filter weak keyword matches. |
Results include path, score, depth, reason (keyword | link | vector), and snippet. Use hits to ground answers or decide what to open next.
Optional cache (not vault)
cache (ClawQL Core, always registered) is ephemeral key/value in this process, LRU-bounded, no Markdown / memory.db. memory_ingest / memory_recall are on by default; set CLAWQL_ENABLE_MEMORY=0 to hide, and use a configured vault to persist. Repo reference: docs/mcp/cache-tool.md.
The server always exposes audit: ephemeral in-process event ring buffer — not the vault and not compliance-grade alone. Use memory_ingest for durable, human-inspectable trails. Repo reference: docs/mcp/enterprise-mcp-tools.md (#89).
When CLAWQL_ENABLE_NOTIFY is set, the server exposes notify: Slack chat.postMessage — not vault storage; use for completion signals alongside memory_ingest when you also want a durable note. Repo reference: docs/mcp/notify-tool.md (#77).
Thorough ingest pattern (recommended structure)
Use insights as the main artifact. A solid long-form layout:
## Summary
One paragraph: what this note captures and why it matters.
## Tags / topics
#grpc #mcp #clawql (Obsidian-compatible hashtags; add domain tags freely)
## Decisions
- Bullet decisions with rationale.
## Commands / env
- `VAR=value` …
- `CLAWQL_MEMORY_INGEST_FILE_ROOTS=/abs/path/to/ClawQL` — optional comma-separated allowlist for **`toolOutputsFile`** (default: process **`cwd`** only). `CLAWQL_MEMORY_INGEST_FILE_MAX_BYTES` (default 10M). `CLAWQL_MEMORY_INGEST_FILE=0` disables path-based reads.
## APIs & references
- Paths, RPC names, links to repo files.
## Risks & caveats
- What breaks, versioning, known grpcurl quirks.
## Follow-ups
- [ ] Next steps
Put raw command output, stack traces, or large JSON in toolOutputs (or, when the text is too large to pass in MCP tool JSON, toolOutputsFile on the server), not inside insights, unless tiny.
Wikilinking strategy
- Link related note titles that already exist or will exist (
wikilinks: ["gRPC README", "mcp-grpc-transport"]).
- Prefer consistent title strings so Obsidian resolves
[[...]] edges predictably.
- For cross-repo concepts, one note can link to ADR, runbook, and feature notes.
Session threading
- Reuse the same
title across a multi-step task and set sessionId per ingest (e.g. session-2026-04-17-grpc) so section headers show lineage.
Append vs new note
- Same topic, new facts: same
title, append: true (default).
- New topic: new
title.
- Full replace (destructive):
append: false—overwrites file body; use sparingly.
Recall-first workflow
memory_recall with a focused query (feature name, error string, component).- Tune
maxDepth if graph context matters (dependencies between notes).
- Summarize relevant hits before editing code or answering from memory alone.
Anti-patterns
- Dumping entire chats into
insights without a Summary—hard to scan later.
wikilinks to titles that do not match vault pages—broken graph edges.- Secrets in any field.
- Relying on custom frontmatter tags via the tool—they are fixed; use hashtags or a Tags line in
insights instead.
Minimal API reminder
memory_ingest: title (required), insights?, conversation?, toolOutputs? (string | string[]), toolOutputsFile? (path on server), wikilinks? (string[]), sessionId?, append? (boolean)
memory_recall: query (required), limit?, maxDepth?, minScore?
Repo details: docs/mcp/mcp-tools.md (section memory_ingest).
For implementation details (SQLite/pgvector, index), see docs/memory/memory-db-hybrid-implementation.md in the repo when relevant.
