wuisabel-gif/MemWhale

Persistent local debugging memory for developers and coding agents.
English README · العربية · Deutsch · README français · 简体中文 README · 繁體中文 README · 한국어 README · 日本語 README · Español · Português (Brasil)
**You hit an error. MemoryWhale remembers you've seen it before, and hands back the fix.**
MemoryWhale records what actually happened while you debug: commands, output,
failures, and the fixes that worked. It stores that evidence in local SQLite so
you and your coding agents can find it after the terminal, SSH connection, or
agent session is gone. It works with 19 coding agents through MCP, with no
account and nothing uploaded.
> On LongMemEval, a public long-term memory benchmark, search puts a
> session holding the answer in its top 5 for **97%** of 470 questions
> (retrieval only, no model).
>
> In a controlled evaluation on project-specific bugs, an agent solved **25%**
> without memory and **96%** with it. These are synthetic tasks that
> demonstrate the mechanism, not a field study; see
> how it was measured.
**MemoryWhale 0.20.0: Agents Learn from Each Other · October 14, 2026.**
The CLI, web UI, and desktop app share product version 0.20.0; the reusable
Rust core is version 0.10.0. See the release notes
for the upgrade guide. Opening a store migrates it from schema 10 to 14.
**Want to contribute?** Start with the Start here issue.
Many tasks need no Rust, including translation reviews and documentation fixes.
Why MemoryWhale
- **Remember what actually happened.** Preserve the command, environment,
output, failure, and lesson—not only a shell-history line.
- **Use one memory across coding agents.** Any compatible stdio MCP client can
read and write the same local memory through mw-mcp.
- **Keep development history local.** MemoryWhale works without an account,
hosted service, or per-token memory bill.
MemoryWhale records development experience, not everything. It is a debugging
memory layer, not an autonomous coding agent, a general-purpose personal memory
system, or a replacement for project documentation.
New in Agent-Native Memory
- **Connect and inspect agents.** Install Claude Code or Rho MCP access,
capture hooks, and memory-use guidance with mw integrate; mw doctor
checks MCP, hooks, and skills independently.
- **Keep provenance explicit.** Schema 10 stores command agents as claude,
rho, or NULL. The display/filter label terminal means terminal/manual
or legacy provenance, not proof that a human ran it. Agent identity is
separate from source type such as command, session, or note.
- **Share a repository, distinguish worktrees.** Canonical repository IDs
group linked worktrees while preserving each worktree root and existing
project tags. Discovery reads local Git metadata, not a remote service.
- **Use local interfaces.** mw-serve provides HTTP MCP at POST /mcp;
mw-serve --api opts into the read-only JSON API. Both use the dashboard's
listener; non-loopback access requires a token.
- **Fetch GitHub context explicitly.** mw github context reads PR
metadata, checks, and reviews through your existing gh login. It prints
bounded, redacted context without checking out code or automatically saving
it to memory. There is no background GitHub sync.
Install
Prebuilt binaries are available for Linux x86_64/aarch64 and macOS: