emilandClaude Opus 4.7 d484de8390 Bump default n_ctx 8192 → 32768
8K filled up after one month of activity (~200 MB state) and the next
`memwalk update` crashed with llama_decode returned 1 because new
tokens didn't fit. Nemotron 3 Nano 4B is trained for 1M context so
32K is well within model capacity and leaves headroom for several
more months of ingest before rotation matters.

Note: existing users need to delete current.memb after upgrading,
since state files are tied to the n_ctx they were created with.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-16 14:10:01 +03:00
2026-05-16 14:10:01 +03:00

memwalk

Walk through your work memory.

A local-first CLI that watches your git activity (and optionally your shell history) and feeds it into a Mamba-based LLM via persistent state. You can then ask in plain English what you were doing last week, why you started that branch, or generate a standup from yesterday's commits — without your data ever leaving the machine.

Built on memba for state persistence and NVIDIA Nemotron-3-Nano-4B (or any other GGUF SSM/hybrid model) for inference.

Status

v0.1 — alpha. Works end-to-end on Linux for the maintainer; APIs and on-disk format may change.

Install

# Until memba is on PyPI, install both editable from local clones:
pip install -e /path/to/Memba
pip install -e /path/to/memwalk

Make sure you have a GGUF Mamba-2 or hybrid model. Recommended:

hf download nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF \
    NVIDIA-Nemotron3-Nano-4B-Q4_K_M.gguf \
    --local-dir ~/.memwalk/models

Quickstart

memwalk init                                  # interactive setup
memwalk update                                # ingest last 30 days of git+bash
memwalk standup                               # auto-generate daily standup
memwalk ask "What was I working on last week?"
memwalk status

What it actually does

memwalk update walks your configured git repos and (optionally) your bash history, formats new events into a readable activity block, and feeds that into the SSM model. The model's hidden state — a fixed ~85 MB blob — is saved to ~/.memwalk/current.memb via memba.

memwalk ask and memwalk standup load that state and query it. The model recalls themes, projects, and trajectory across processes and reboots.

Use from an agent (MCP)

memwalk ships an MCP server so Claude Code / opencode / any MCP-aware agent can query your memory as native tools.

memwalk mcp     # starts a stdio MCP server

Tools exposed: ask(question), standup(), status(), update(). The Session loads lazily on the first call that needs it, then stays in memory — first call ~2 s, subsequent calls <500 ms.

Configure Claude Code

Easiest way (Claude Code CLI):

claude mcp add memwalk -- memwalk mcp

Or by hand, in ~/.claude/mcp_servers.json (path may vary by version):

{
  "mcpServers": {
    "memwalk": {
      "command": "memwalk",
      "args": ["mcp"]
    }
  }
}

Restart Claude Code. Tools appear as mcp__memwalk__ask, mcp__memwalk__standup, etc.

Configure opencode

opencode uses its own MCP block in opencode.json:

{
  "mcpServers": {
    "memwalk": { "command": "memwalk", "args": ["mcp"] }
  }
}

Layout

~/.memwalk/
├── config.toml
├── current.memb              ← rolling state
├── last_update.txt
└── snapshots/
    └── 2026-05-16.memb       ← daily snapshot before each update

License

MIT.

S
Description
No description provided
Readme
117 KiB
Languages
Python 100%