Mneme

What is Mneme?

Mneme is a vault-native memory server for Claude Code and other MCP clients. It stores conversation memory as plain Markdown files in a directory (vault), making the data git-diffable, grepable, and editable.

How to use Mneme?

Install the Claude Code plugin with pipx install mneme-cc-plugin && mneme install. Alternatively, install the npm package mneme-mcp-server for use with any MCP client. No cloud account is required; the vault is a local directory of markdown files.

Key features of Mneme

• Markdown as ground truth, no opaque binary storage
• No LLM call on the critical path; deterministic Stop capture
• Hybrid retrieval with FTS5 BM25 and opt-in local dense embeddings
• Token-efficient architecture with adaptive context budget
• Built-in <private> tag redaction with SHA256 audit log
• Temporal claim lifecycle with valid-from/to, supersedes, and provenance time-travel

Use cases of Mneme

• Persistent, editable memory for Claude Code sessions
• Team memory via self-hosted git sync with redaction-before-share
• Code memory through procedural AGENTS.md parsing and fix-trajectory recording
• Project and code graph analysis with tree-sitter extraction
• Agent autonomy with policy-graduated edits and tamper-evident audit log

FAQ from Mneme

What makes Mneme different from other memory tools?

Mneme stores all memory as plain Markdown files instead of opaque SQLite blobs. It never calls an LLM on session end (summarization is deterministic and zero-LLM by default), and it runs entirely locally with no cloud account required.

Does Mneme require Obsidian?

No. A vault is simply a directory of plain .md files. You can use grep, git, VS Code, or any text editor. Obsidian is fully optional; if you already use it, you can point it at the same directory for rendered notes and graph-view navigation.

Is Mneme local-first?

Yes. All memory data, indexes, staging, and audit logs stay in a .mneme directory within the vault. No outbound network calls occur except for the opt-in compression LLM and optional local Neo4j instance.

Does Mneme use an LLM for summarization?

No, the default session summary is deterministic and zero-LLM. LLM-based compression is available as an opt-in feature with a monthly cost-cap ledger, but never runs on the Stop hook or critical path.

What are the runtime dependencies of Mneme?

The core packages (mneme-core, mneme-cc-plugin, mneme-mcp-server) require Python or Node.js. Optional dependencies include sentence-transformers for dense retrieval and Docker + Neo4j for knowledge-graph temporal enrichment. No cloud infrastructure is required.