MengBot
MEMORY ATLAS · MEMORY-ETERNAL
中文

Give your AI
a memory it can trust

Conversations are distilled into knowledge cards, reviewed before they enter the vault, and measured with a built-in token ledger. One local SQLite vault shared by DSH / Claude Code / Codex / Cursor / any MCP client.

GitHubLive demoQuick start
58 starsMITv0.16.0local-firstMCPself-hosted
memory-eternal demo: knowledge cards, review inbox, token ledger, knowledge graph
MEMORY-ETERNAL · 14S TOURREVIEW · LEDGER · GRAPH
AT A GLANCE01 / 06
01
1
local SQLite vault
cards / review queue / audit log / ledger
02
6+
agent entry points
DSH · Claude Code · Codex · Cursor · MCP · TUI
03
0
cloud dependencies
data stays on your machine, works offline
04
MIT
open-source license
self-host it, build on it
WHY IT IS DIFFERENT02 / 06

Not another vector store — an auditable memory layer

The hard part is not storing memories. It is making them trustworthy, making their cost visible, and making them usable by every agent. Those three decide whether memory survives daily use.

REVIEW FIRST01

Review first

New cards land in a pending queue and only enter the main vault after human approval. Every status change is appended to an immutable audit log — who approved what, and when.

TOKEN LEDGER02

Token ledger

Recall injection, distillation calls and usefulness are recorded. The README even explains when memory does *not* save tokens — measure first, optimize later.

LOCAL-FIRST03

Local-first · multi-host

SQLite + Markdown on your own disk, no third-party memory framework, no account. DSH, Claude Code, Codex, Cursor and any MCP client share the same vault.

PRODUCT TOUR03 / 06

Real interfaces, not mockups

Open live demo
T-01Click to enlarge
Knowledge cards
Knowledge cardsSearch, type filters, daily resurface and auto-capture badges.
T-02Click to enlarge
Review inbox
Review inboxPending queue plus a verdict pane: approve, reject, bulk-process.
T-03Click to enlarge
Knowledge graph
Knowledge graphKind colours, tag links and cross-vault aggregation.
L-01TOKEN LEDGER

Does memory actually save tokens?

Most memory products claim they save tokens. memory-eternal ships a ledger that records recall injection, distillation calls and usefulness — measure first, conclude later.

net benefit ≈ avoided context cost − (injected characters + distillation calls × price)
  • SAVESSaves: multi-session projects, many past decisions, avoiding repeated mistakes
  • NO GAINDoes not save: one-off tasks, low hit rate, distillation nobody reuses
  • TUNEMeasure yours with the ledger; if usefulness stays under ~60%, tune it down or turn memory off
Memory ledger: recall injection, distillation calls, estimated tokens and recall usefulness

RUNTIME STATUS → MEMORY LEDGER

QUICK START04 / 06

Up and running in minutes

Auto-capture, the review inbox and the MCP tools are ready after install. See the GitHub repo for everything else.

DeepSeek HarnessSRC
dsh plugin --profile desktop add memory-eternal
# reopen DSH → sidebar Memory
Claude Code / Codex / CursorSRC
npm i -g memory-eternal
dsh-memory connect claude
# or: codex / cursor
Browser / terminalSRC
dsh-memory open
npx -y -p memory-eternal dsh-memory tui
Full install / platform support / CLI reference: github.com/EternalNight996/memory-eternal
O-01OPEN SOURCE

MIT licensed, your data stays yours

No cloud, no accounts. Run it locally, or host it on your own server for intranet access.

Deployment guideFull manual
O-02SELF-HOSTED

One command on your own server

git clone https://github.com/EternalNight996/memory-eternal.git
cd memory-eternal
docker compose -f deploy/docker-compose.yml up -d --build

A Caddy (TLS + Basic Auth) example is included; it binds 127.0.0.1 by default — add a reverse proxy with auth for remote access.

NEXT STEP / 下一步

If it saved you from explaining the same context again

A star helps more people who are tired of "my AI forgot again" find it. Or start with the live demo.

Star on GitHubOpen live demo