intermem

intermem

Memory synthesis for Claude Code.

What this does

Claude Code's auto-memory accumulates facts over time, but they live in a flat file that grows without bound. Some of those facts are stable and well-confirmed; others are speculative or stale. intermem graduates the stable ones into curated reference documents (AGENTS.md, CLAUDE.md) and prunes the source.

The pipeline: scan memory files → detect stability (same fact confirmed across multiple sessions via SHA-256 content hashing) → validate citations and paths → deduplicate against existing docs → present for interactive approval → promote to target document + prune from source. A WAL-style journal ensures atomicity: if the process crashes mid-promotion, it recovers cleanly.

Installation

First, add the interagency marketplace (one-time setup):

/plugin marketplace add mistakeknot/interagency-marketplace

Then install the plugin:

/plugin install intermem

Usage

Run the synthesis pipeline:

/intermem:synthesize

Validate memory health without promoting:

/intermem:validate

Architecture

intermem/            Python library (scan, stability, validation, dedup, promotion)
skills/              synthesize, validate
.intermem/           Per-project state (metadata.db in SQLite)

Design decisions

• Confidence scoring: base 0.5, +0.3 for valid citations, -0.4 for broken citations, +0.2 for 5+ snapshot appearances. Below 0.3 is considered stale.
• No hooks (Clavain hook budget constraint): synthesis is always manually triggered
• No MCP server: skills are sufficient for this workflow
• .intermem/ must be in .gitignore (local state, not shared)