Session-start recall
The SessionStart hook queries Remnic with your
project context and injects the result before Claude sees the first
message.
Claude Code integration
No memories appearing in
Claude Code that remembers your repo.
@remnic/plugin-claude-code plugs Remnic into Claude Code
through its session hooks. Sessions start with your project context,
each prompt recalls relevant memories, and every edit is observed for
later extraction. Recall is wired into the hooks, so the agent does not
have to ask for it.
What you get
Per-prompt recall
Every user message fires UserPromptSubmit. Remnic runs
a minimal-mode recall with the prompt as the query and injects the top
matches inline.
Edit-time observe
The PostToolUse hook runs after Write, Edit, and
MultiEdit tools. It sends the change to Remnic in the background for
extraction. It never blocks Claude.
MCP tools and skills
The plugin registers the Remnic MCP tool surface plus memory skills and a memory-review agent, so recall, store, and entity lookup are also available on demand.
Install
Three steps, each writing to a different place. Start the Remnic daemon first, then wire Claude Code to it.
# 1. Install Remnic and start the daemon
npm install -g @remnic/cli && remnic daemon install
# 2. Mint a Claude Code token and record the connector
remnic connectors install claude-code
# 3. Install the plugin, then load it through Claude Code
npm install -g @remnic/plugin-claude-code
Step 2 writes Remnic-side connector state and a bearer token. It does
not edit Claude Code's config. Between steps 2 and 3, paste the
.mcp.json block from the package README into Claude Code's
MCP config, pointing it at http://localhost:4318/mcp with
that token. Load the plugin through Claude Code's plugin loader, then
restart Claude Code so the hooks fire.
Manual wiring: the exact hook list, the
.mcp.json block, and the plugin-load steps live in the
package README.
Memory skills
The plugin ships skills you can invoke directly in a Claude Code chat.
| Command | Description |
|---|---|
/remnic-remember <text> | Store a memory explicitly ("I prefer tabs over spaces"). |
/remnic-recall <query> | Recall memories ("what do I know about auth?"). |
/remnic-search <query> | Full-text search across all memories. |
/remnic-entities | List tracked entities: people, projects, tools. |
/remnic-status | Show Remnic connection status and memory stats. |
Cross-project memory in action
The killer benefit is that memories from one project surface in another without you doing anything:
- You fix a tricky Zod
optional().nullable()bug in Project A. - Remnic extracts: "Zod optional fields must use
.optional().nullable(), not just.optional()". - Later, in Project B, you start writing a Zod schema.
- The
UserPromptSubmithook recalls the memory. Claude sees it inadditionalContextbefore you ask. - The bug never happens again.
Project-scoped and procedural memory
The session-start hook detects your git repository and scopes memory to that project, so what Claude learns in one repo does not surface in another. Global facts, such as a library's behavior or your own preferences, still carry across projects. Remnic also captures procedural memory: multi-step runbooks distilled from your work, on by default outside the conservative preset. More detail is on the coding agents page.
Troubleshooting
No memories appearing in additionalContext
Check ~/.remnic/logs/remnic-session-recall.log and
~/.remnic/logs/remnic-user-prompt-recall.log. Prompts
shorter than 4 words are skipped. Confirm the daemon is up with
remnic daemon status.
Per-turn recall feels slow
UserPromptSubmit uses minimal mode with a 20s timeout.
If it's still too slow, raise the PROMPT_WORD_COUNT
threshold in the hook script or reduce topK from 8 to 4.
Token not found
If Claude Code is launched from the Dock or Spotlight rather than a
terminal, it doesn't inherit shell variables. Either add the export to
~/.zprofile or run
launchctl setenv OPENCLAW_REMNIC_ACCESS_TOKEN "your-token".