MCP 服务器

coldframe

Local-first memory for your AI agent. One SQLite file you own — offline, no API key. Plugs straight into Claude Code.

README

<h1 align="center">coldframe</h1>

<p align="center"> <b>Local-first memory for your AI agent.</b><br> One SQLite file you own — offline, no API key. Plugs straight into Claude Code. </p>

Your agent forgets everything between sessions. coldframe gives it a memory that lives in one local file (~/.cold-frame/memory.db) — with full-text + vector search, version history, automatic forgetting, and the ability to rewind what was true when. No server, no account, no key.

Quickstart

In Claude Code — one install, then it remembers automatically:

claude plugin marketplace add coldzero94/cold-frame
claude plugin install coldframe

Recall is injected at the start of every session; durable facts you state are captured as you work (using the Claude you already pay for — no extra key or cost). Nothing to configure.

On the command line — offline, instant:

cold-frame add "I prefer dark roast coffee."
cold-frame search "what coffee do I like?"
cold-frame doctor          # health: counts, integrity, embedder

(Install: uv tool install "cold-frame[mcp]" · Homebrew · or a single binary — see packaging/. A PyPI release is pending name clearance.)

Why coldframe

🗃️ One file you own — everything is ~/.cold-frame/memory.db. Copy it to back up, delete it to forget. No server, no account.
🔌 Offline by default — add → search works with zero keys, zero network.
🤖 Automatic in Claude Code — recalls what matters, captures durable facts as you work — and the engine keeps it lean instead of hoarding everything.
⏪ Rewindable belief — facts are corrected and superseded, never silently overwritten. Ask "what did I believe back in March?" and get the answer as of that date.

How the memory works

Recall — a SessionStart hook injects your strongest durable memories; a UserPromptSubmit hook adds ones relevant to the current prompt (gated so it adds signal, not noise).
Capture — Claude itself extracts durable facts and calls add_memory during your session (a bundled skill), with a keyless deterministic backstop so nothing is missed; dedup merges them.
Stays lean — every ~20 facts a consolidation pass decays, rolls up, and archives the weakest (per-scope caps), so the active set never grows without bound. Forgetting archives — nothing is truly deleted, and obvious secrets (API keys, tokens, private keys) are blocked before they ever touch disk — though PII redaction and at-rest encryption are not yet automatic (planned).
Per-project + global — facts are tagged by git project; clear personal facts go to a global tier recalled everywhere.

It's a standard stdio MCP server, so it works with any MCP client — not just Claude Code.

Use it as a library

from datetime import UTC, datetime
from cold_frame import Memory, Scope

mem = Memory()                                   # ~/.cold-frame/memory.db, offline
mem.add("I switched jobs to Anthropic in 2026.", scope=Scope(user_id="coby"))
print(mem.search("where do I work?").hits[0].note.content)

past = mem.search("where do I work?", as_of=datetime(2026, 3, 1, tzinfo=UTC))  # rewind

See and edit your memory anytime: cold-frame ui — a local dashboard at 127.0.0.1:27182 where you can browse facts and pin / correct / forget them (writes are CSRF-guarded and localhost-only).

Install & distribution — packaging/ (plugin, Homebrew, standalone binary)
Design & spec — docs/ (the spec, decisions, and the analysis of mem0 / Letta / Zep-Graphiti / Cognee / MemOS / A-MEM / LangMem that informed it)
Back up / move — cold-frame export backup.db then import it elsewhere.

Status

The engine, CLI, MCP server, Claude Code plugin, secret-blocking + grep-verified hard-purge, and a local web UI are built and tested end-to-end on a fully offline gate (ruff + mypy --strict + ~340 deterministic mock-LLM tests green). Planned: full PII redaction + crypto-shred purge, a write web UI, event-log replay import, and a PyPI release (the name is pending trademark/registry clearance).

Apache-2.0.