coldframe

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

_{Your memory as a banked fire — warmth is belief, the cold is forgetting. Pinned facts sit under glass; the weak gutter out.}

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 "dark roast coffee"
cold-frame list            # each memory with its strength band
cold-frame doctor          # health: counts, integrity, embedder

(Install: brew install coldzero94/coldframe/cold-frame (Apple Silicon macOS + Linux x86_64), or download the self-contained binary for your platform from the latest release — CLI + MCP in one file, no Python needed. The optional [local-llm] (semantic recall) extra isn't in the binary; install from source for that — see examples/.)

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• ⏪ Rewindable belief — facts are corrected and superseded, never silently overwritten.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.
• 🧭 Conflict resolution — supersession is always deterministic (code decides by valid_at, never the LLM). Detecting that two facts contradict needs a model: the cold-frame worker (and COLD_FRAME_LLM=claude, using your Claude session — no API key) turn it on; the zero-key offline default still does exact/near-duplicate merging and honors explicit correct/supersede.

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. PII redaction (email/phone/card/ssn) is available opt-in (add --redact-pii, or Memory(pii_redact=…)) — off by default so a personal store keeps your own contact facts. For at-rest protection, the DB is a single local file — put it on an OS-encrypted disk (FileVault / LUKS), which covers the stolen-laptop threat without a second key to manage.
• Per-project + global — facts are tagged by git project; clear personal facts go to a global tier recalled everywhere.
• Semantic recall (opt-in) — the zero-config default recall is lexical (offline HashEmbedder, no download). For semantic recall, install from source with the [local-llm] extra and set COLD_FRAME_EMBEDDER=local (a small local bge-small model, downloaded once — nothing leaves the machine). After switching on an existing DB, run cold-frame reembed to re-index.

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
me = Scope(user_id="coby")
mem.add("I switched jobs to Anthropic in 2026.", scope=me)
print(mem.search("where do I work?", scope=me).hits[0].note.content)

past = mem.search("where do I work?", scope=me, 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).

More

• 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, deterministic pre-disk secret-blocking + grep-verified hard-purge, opt-in PII redaction, an admission confidence-gate + opt-in consent hold, idempotent event-log replay import, and a local web UI are built and tested end-to-end on a fully offline gate (ruff + mypy --strict + ~420 deterministic mock-LLM tests green on a 3.12/3.13 CI matrix). The automatic recall + capture loop is verified end-to-end against real Claude Code. Distributed as a self-contained binary via Homebrew + GitHub Releases (not PyPI).

Apache-2.0.