tunc-clm 0.2.0

Parser and v3.0 trim-aware validator for CLM (Claude Memory Format)

tunc-clm (Python)

Parser and v3.0 trim-aware validator for the Claude Memory Format. Pure Python, no runtime dependencies, type-hinted.

pip install tunc-clm

from clm import Document, validate_v3, validate_v3_with_filesystem

doc = Document.parse(open("MANIFESTO.clm").read())
assert str(doc) == open("MANIFESTO.clm").read()  # round-trip byte-identical

report = validate_v3(doc)
print(f"{len(report.errors)} errors, {len(report.warnings)} warnings")

# Filesystem-aware variant cross-checks the sibling archive
report = validate_v3_with_filesystem(doc, base_dir=".")

CLI

clm validate path/to/file.clm

What CLM is for

CLM/3.0 is a write-ahead log for multi-session AI handoff threads. Four axes; this package implements two of them:

1. Read retrieval — prose summary wins this axis; don't pick CLM for one-shot Q&A.
2. Write cost — CLM wins by 4.2× at 100 sessions, 12.2× at 500. (See experiments/v3/RESULTS-compounding-cost.md.)
3. Audit integrity — verbatim preservation, signed deltas. CLM by ritual.
4. Tooling — parser round-trip + validator. This package implements axes 3 and 4 in Python; the Rust and TypeScript packages do the same.

See the main README for the full positioning, including links to the four review rounds where each marketing claim was iteratively narrowed by @copyleftdev's empirical work.

What this validator checks

Per SPEC.clm validation.posture.v3.0:

• Header declarations: trim.mode, trim.config, archive.mode, archive.path
• Trim-config grammar: keys / duplicates / missing values / unknown keys
• Lifecycle states A/B/C; declared offload via (last X of Y archived) form
• Sentinel placement (BEFORE entries, not after)
• Per-entry shape (ROLL.CALL needs · YYYY-MM-DD ·; DREAM.LOG needs | YYYY-MM-DD |); malformed quarantined
• Cross-doc sentinel symmetry (live ↔ archive)
• Archive structural check (must contain trim ARCHIVE sections, not just any file)
• Filesystem-aware variant (state.B → warning, state.C → error)
• Chained archive validation (warnings propagate from archive into live report)

Mirror of the Rust reference

This package mirrors clm-rs and is paired with clm-js. All three implementations validate the same set of behaviors against the same canonical artifacts (MANIFESTO.clm, SPEC.clm, the experiments/v3/ bench docs). 117 tests across all three.

Known limitations (parity with clm-rs v0.2.0)

Three follow-ups from Codex round-8 review apply equally:

1. [DECISIONS.ARCHIVE] cross-doc sentinel check is missing (symmetric gap to [ROLL.CALL.ARCHIVE] / [DREAM.LOG.ARCHIVE]).
2. Malformed lines in decisions.live aren't quarantined (counted toward overflow).
3. Generator (experiments/v3/gen_50_session.py) emits nonsense metadata for depths ≤ 5.

License

MIT.

Audit thread

CLM is itself an append-only audit-thread format. The thread for MANIFESTO.clm and SPEC.clm is at https://github.com/TUNC-AI/tunc-clm — read it like any CLM file: open, sign, append.