Adaptive cognitive memory

A Claude Code plugin that gives Claude a persistent cognitive memory: one central, local SQLite database shared by every project, with local semantic search, budgeted hook recall and explicit, curated capture.

Every project Claude works in reads and writes the same ${CLAUDE_PLUGIN_DATA}/memory.sqlite. Memories carry a scope — project (visible only in that project) or user (visible everywhere) — and a type: semantic (facts), procedural (how-tos), episodic (session events) or pattern (lessons and preferences). All memories are stored in English as the canonical retrieval language; a multilingual local embedding model and the recall questions each memory carries, written in English and in the user's language, let prompts in other languages find them.

How it works

• Recall is automatic. Hooks inject one-line abstracts under a token budget at session start, on every user prompt, before tool calls and after a failed Bash command; Claude expands a memory's full body with memory_get.
• Capture is explicit. Nothing is auto-saved. Claude stores durable knowledge with the memory_* MCP tools, guided by the bundled memory skill. Claude notices a correction or a remember request itself, in any language; the hooks only point at a failing command that later passed.
• Claude curates, the server guards. A store that finds related memories writes nothing and returns them; Claude decides to supersede, update, skip or add. Superseded memories stay readable but never reach recall.
• Nothing ever leaves the machine. No outbound request carries memory content, the MCP server talks stdio only, and hooks make zero LLM calls. The one network operation is the one-time embedding model download. Secrets are redacted at the persist boundary and again on export.
• Degraded is never silent. A missing embedding model or sqlite-vec extension falls back to FTS5 keyword search with a visible warning; a native module that does not load or a database from a newer plugin turns memory off with one session-start notice; /memory-doctor names the broken piece and its fix.

Requirements: Claude Code and Node.js ≥ 22. To start, follow Install.

Uninstalling the plugin from the last scope it is installed at deletes every memory. Export them first: Uninstall.

Documentation

Using the plugin

• Install and uninstall: install from the marketplace or a local checkout, the embedding model download, keeping your memories when you uninstall
• Recall: what Claude sees, when each hook recalls, turning recall off, branch-aware session start, recall after a failed Bash command
• Capture: the two-phase store, capture moments, the Stop checkpoint and what the hooks print
• MCP tools and skills: every memory_* tool with its arguments and result, and the memory-curate skill
• Commands and CLI: /memory-doctor, /memory-recall, /memory-statusline, export, import, re-embedding and the doctor checks
• Statusline badge: the opt-in badge and the line for your statusline script
• Configuration: plugin options, environment variables and native auto memory

Developing the plugin

• Development: where everything lives, checks, the test suite, performance gates and the real-session eval
• Storage: the database tables, project ids, concurrency and schema versions
• Contributing