MnemoDB

A structured, portable, auditable memory format for AI agents — a database whose files you can read.

claude mcp add mnemodb -- npx -y @mnemodb/[email protected]

Repo, spec, and security audit: https://github.com/mnemodb/mnemodb

What it is

A plain-text, Markdown-compatible file format (.mem.md) for the persistent memory of AI agents, plus the engine and tooling around it. A memory is a Markdown heading with a one-line metadata span:

## decision: We use PostgreSQL LISTEN/NOTIFY for cache invalidation, not Redis
`mnemo c4d1 | src: user | conf: high | supersedes: 99e0`

Redis was dropped to cut infra count. Revisit if we exceed ~500 notifications/s.

It renders in any Markdown viewer, diffs cleanly in git, and a program reads every field unambiguously. Every existing CLAUDE.md / AGENTS.md is already a valid (untyped) MnemoDB file — adoption requires renaming and migrating nothing.

Why — vs the memory you may already have

AI agents already have some memory: CLAUDE.md loads a flat file into context, and Claude's memory tool gives the model a folder to organize however it likes. Both work. MnemoDB isn't a replacement — it's the structure and discipline they don't provide.

"Give the model a folder and trust it" is simple and often good enough. MnemoDB is the opposite bet — a defined format with real guarantees:

• Lifecycle, not sprawl — typed entries with TTLs, supersession, and a compact pass that archives what's expired or replaced. A flat memory file only grows and rots; MnemoDB forgets on purpose, auditably.
• Provenance & trust — every memory records its source (user / agent / tool). Tool-sourced content is flagged untrusted and can never supersede a human instruction, so a poisoned memory can't hijack future sessions.
• Portable — the same files work across Claude Code, Cursor, and any MCP client. Not locked to one vendor.
• Auditable — readable Markdown in your git repo; git diff shows exactly what your agent learned, and mnemo doctor lints for staleness, contradictions, and damage.

Worth it when you want the agent to maintain memory itself across many sessions without decay, you work across more than one AI tool, or you need to audit and trust what's in memory. Overkill when a short hand-curated CLAUDE.md already does the job — we'd rather say so up front.

How to use it

npx @mnemodb/cli init                      # create a store (or migrate CLAUDE.md)
claude mcp add mnemodb -- npx -y @mnemodb/[email protected]   # give your agent the tools
npx @mnemodb/cli list                      # inspect anytime — it's just files

Then add a short instruction to CLAUDE.md so the agent actually calls the tools. Full walkthrough: https://github.com/mnemodb/mnemodb/blob/main/docs/USAGE.md

Packages

• mnemodb (this package) — umbrella; re-exports @mnemodb/core and bundles the spec
• @mnemodb/core — parse, serialize, index, resolve, lifecycle, merge, validate (zero deps)
• @mnemodb/cli — the mnemo command line
• @mnemodb/mcp — the memory engine as an MCP server for Claude Code, Cursor, any MCP client

Status

v0.1 — early. Honest feedback wanted, not stars. Security-hardened by an adversarial audit; 46 tests enforced as a ship gate on every release.

License: Apache-2.0 (code), CC BY 4.0 (spec).