memory-eternal plugin
Local memory for agents with a review queue: SQLite, an audit log and a token ledger, one vault for Claude Code, Codex and Cursor
- 180 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Agent sessions are condensed by a model and saved as cards that contain your project context
- Approved memory automatically enters the context of every session
- The News page requests public RSS, GitHub, Hacker News and arXiv on demand
Install
- Works in
Manual install
npm i -g memory-eternal && dsh-memory connect claudeFor Codex and Cursor replace `claude` with `codex` or `cursor`; for DSH `dsh plugin --profile desktop add memory-eternal`.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Agent sessions are condensed by a model and saved as cards that contain your project context
- Approved memory automatically enters the context of every session
- The News page requests public RSS, GitHub, Hacker News and arXiv on demand
This is third-party code. Review the repository files before installing.
What it does
memory-eternal keeps long-term AI memory in a single SQLite vault with a Markdown mirror on your disk, with no cloud or account. A session is condensed by a distillation step through an LLM, and on failure a raw card stays so nothing is lost. New cards go into a review queue: only a human approves or rejects in the web UI or terminal, and every change is written to an immutable log. Memory can be edited by hand, and an agent can submit a revised copy through `memory_revise`: an approved card is overwritten only after your approval. Approved cards can be promoted to core memory that every session gets. Search uses FTS5 with BM25, and the token ledger shows how much went into recall and how often it helped. For Claude Code, Codex and Cursor access goes over MCP, and for DSH it installs as a plugin.
Who it is for. People who run projects with several agents and want shared long-term memory where nothing enters context without a human review.
Good fit when
- One set of knowledge must be shared by Claude Code, Codex and Cursor
- Agent sessions must be turned into cards and reviewed before entering memory
- You need to see how many tokens recall spent and whether it helped
- An agent proposed a correction to an entry and it must be approved by hand
Not a fit when
- Fully automatic memory with no human review is needed: the approval queue is mandatory
- Not a single LLM call is acceptable: the distillation step uses a model
- Cloud sync between machines is needed: the data lives only locally
Example request
Connect the shared memory from memory-eternal and show which cards are waiting for my approval.Limitations
Node.js for the `memory-eternal` npm package and a connected model for the distillation step are required. Data is stored locally in SQLite and Markdown, and the News page reads public RSS and Hacker News on demand (titles, excerpts and links). New cards do not reach the main store before approval. Parts of the README are in Chinese and English, and clients like MengBot, OpenClaw and Hermes are described in `docs/INTEGRATIONS.md`. The latest commit in the repository is from November 2026.
How to disable. Remove the client entry from the Claude Code, Codex or Cursor configuration and uninstall the package with `npm uninstall -g memory-eternal`. The SQLite vault and the Markdown mirror sit in the data directory and are deleted by hand.
FAQ
How do I connect it to Claude Code?
With `npm i -g memory-eternal` and `dsh-memory connect claude`. For Codex and Cursor replace `claude` with `codex` or `cursor`, and the vault is shared.
How do I install for DSH?
With `dsh plugin --profile desktop add memory-eternal`, then reopen DSH: a Memory section appears in the sidebar.
Who approves new cards?
Only a human: in the `dsh-memory open` web UI or the `dsh-memory tui` terminal. The main store holds only approved cards.
What does the token ledger show?
How many tokens recall returned into context, how many distillation calls ran and how often recall turned out useful.
Does data go to the cloud?
The vault is local: SQLite and Markdown on your disk with no account. Exceptions: the distillation step goes through an LLM and the News page requests public feeds.
Discussion
No questions or reviews yet.
Related
PluginsLocal memory for Claude Code, Codex and Claude Desktop: one Markdown folder and one MCP server shared by all.
Long-term memory for AI agents: Markdown files as the source of truth, a SQLite index as a cache, shared by Claude Code, Codex and others
One local memory shared by Claude Code, Codex and OpenClaw: tell one agent, another knows tomorrow, no cloud or keys.
Local memory for an agent: hooks record facts from conversations and the MCP server brings them back in later sessions from one SQLite file
Local persistent memory for coding agents: hooks record every session, and Claude Code, Codex or Cursor recall past decisions through 54 MCP tools
Long-term agent memory in local Markdown files with a link graph, semantic search and MCP access
README in short
The README states the goal and lists the principles: review before memory, manual editing, a token ledger, local storage and core memory. A quick start for DSH, MCP clients and the browser and a data flow diagram follow.