size | MACHINE-MEMORY by JFA

Usage

machine-memory size <content> (--local|--remote)
machine-memory size --from-file <path> (--local|--remote)

Embedding budget

size reports whether a prospective memory would pass the embedding budget without touching the database. It composes the embedding text exactly like an add would (content plus tags/context lines plus fixed type/status/certainty lines) and reports both limits independently:

Every memory’s composed embedding text must stay below 512 BGE tokens, and its conservative byte estimate (UTF-8 bytes + 2) must be at most 512. Either limit rejects the write before anything is stored.

Use these commands to preflight without writing:

machine-memory size "<text>" --remote
machine-memory add "<text>" --dry-run --remote
machine-memory update <id> "<text>" --dry-run --remote

Pass --token-report to a real add or update to include the per-part token breakdown. When a memory exceeds the budget, the error states the exact byte or token deficit and suggests a mechanical trim.

{
  "source": "tokenizer",
  "total_tokens": 16,
  "bytes_estimate": 72,
  "limits": {
    "tokens": { "value": 16, "limit": "below 512", "pass": true },
    "bytes_estimate": { "value": 72, "limit": 512, "pass": true }
  },
  "within_budget": true,
  "binding_limit": null,
  "over_by_tokens": 0,
  "over_by_bytes": 0,
  "parts": [{ "part": "content", "tokens": 4 }]
}

Exit code

The command exits 1 when either limit fails, so scripts can preflight writes:

machine-memory size "$(cat long-note.txt)" || echo "would be rejected"

When over budget, the report includes largest_part and a mechanically derived trimmed_suggestion — an exact UTF-8-safe truncation when the byte+2 estimate binds, or an approximation labeled as such when tokens bind.

Options

--tags, --context, --type, and --certainty are honored so the composed text matches the real write. --from-file reads content from a file. The backend flag is accepted for invocation symmetry but no database is accessed.