Retrain without forgetting (and when to forget)

DLM's retention story: edits to a .dlm add new training signal while previous versions of the document stay in a replay corpus. This recipe walks through the default "additive" retrain, when to deliberately forget, and how to audit what the model has seen.

The default: additive retrain

# v0001 — initial training
$ uv run dlm train tutor.dlm
trained: v0001 (20 steps, seed=42, determinism=best_effort)
adapter: ~/.dlm/store/01KPM…/adapter/versions/v0001
log:     ~/.dlm/store/01KPM…/logs/train-000001-…jsonl

# Edit tutor.dlm — add more Q&A, fix a typo
$ $EDITOR tutor.dlm

# v0002 — retrain
$ uv run dlm train tutor.dlm
trained: v0002 (20 steps, seed=42, determinism=best_effort)
adapter: ~/.dlm/store/01KPM…/adapter/versions/v0002
log:     ~/.dlm/store/01KPM…/logs/train-000002-…jsonl

What dlm train does on retrain:

1. Loads the previous manifest.json; diffs this run's section hashes against the recorded content_hashes.
2. Appends the new/changed sections to the replay corpus under replay/corpus.zst (Sprint 08).
3. Builds the training set as current sections + recency-weighted sample from replay. Prior sections don't disappear — they keep showing up proportional to how recent they are.
4. Commits the new adapter as v0002; adapter/current.txt flips atomically (Sprint 09's two-phase commit).

The adapter from v0002 still remembers v0001's Q&A pairs because the training data for v0002 included them (via replay).

Inspecting what the model has seen

$ uv run dlm show tutor.dlm
tutor.dlm
  dlm_id:         01KPM618S78XK668EX0TFEWAJY
  base_model:     qwen2.5-coder-1.5b (revision abcdef1)
  store:          ~/.dlm/store/01KPM…  (12.4 KB)
  adapter:        v0002
  training runs:  2 — last 2026-04-19T18:30:14
  exports:        0

Pretty-print summary; use --json for the full record including content_hashes, replay_size_bytes, per-run TrainingRunSummary entries, and pinned_versions:

$ uv run dlm show tutor.dlm --json | jq '.training_runs'
2
$ uv run dlm show tutor.dlm --json | jq 'keys'
[
  "adapter_version", "base_model", "base_model_revision",
  "content_hashes", "dlm_id", "exports", "has_adapter_current",
  "last_trained_at", "orphaned", "path", "pinned_versions",
  "replay_size_bytes", "source_path", "total_size_bytes",
  "training_runs"
]

Intentional forgetting: --fresh

Sometimes you want the model to unlearn. Maybe the v0001 Q&A was wrong, and replay-weighted retraining keeps reinforcing it.

$ uv run dlm train tutor.dlm --fresh

--fresh wipes the replay corpus AND the optimizer state; the next run trains only on the current document's sections. The adapter version still increments (v0003), and the manifest records that this run was a fresh start.

Use sparingly — fresh training loses every prior training signal.

Intentional pruning: edit the replay corpus

The replay corpus is a zstd-framed file on disk at <store>/replay/corpus.zst + <store>/replay/index.json. There is no shipped CLI for surgical pruning yet; drop into Python via dlm.replay.store.ReplayStore.at(corpus_path, index_path) to read or rewrite it. A first-class dlm replay prune command is on the Phase 4 roadmap.

What dlm.lock records

After each successful run, dlm.lock records:

• last_run_id — incremented each run
• dlm_sha256 — hash of the source .dlm at this run
• pinned_versions — the tuple that produced this adapter

See Determinism for the full field list and mismatch policy.

Caveats

• Replay is lossy on eviction. The corpus has a soft size cap (Sprint 08) and evicts the oldest, lowest-weight entries when it grows. Everything still in the corpus trains; evictions are lost.
• --fresh doesn't forget the base model. The base is whatever base_model: says in the frontmatter. Its pretraining data lives in its weights, not in your replay corpus.