Falln87/clerk-memory
CLERK paper draft v0.1: method, benchmark spec, protocol; results tables pending observed runs
CLERK project README: method summary, novelty statement, repro instructions
CLERK end-to-end reproducible pipeline (CPU smoke tests -> data -> LoRA SFT -> both evals)
eval_locomo: fix choice_logprobs continuation scoring
eval_locomo: clean rewrite (working choice log-likelihood scorer, consistent templates, single model loader)
CLERK LoCoMo-MC10 zero-shot eval: incremental ledger build over LoCoMo sessions, MC log-likelihood scoring, full-context + prompted-ledger + summary-RAG baselines
generator: fix gold_ops_for_events double-application bug; single-state implementation
tests_smoke: clean rewrite of invariant checks (gold ops executable, budget, round-trip, QA consistency)
eval_synthetic: initialize stale-error counters in Runner.__init__
CLERK pinned dependencies
CLERK smoke tests: gold ops executable, ledger round-trip, budget enforcement, QA supervision consistency
eval_synthetic: full clean rewrite (consistent generate contract, main, serialization helper)
CLERK synthetic eval: clerk rollout + prompted-ledger/full-context/sliding-window/RAG baselines, per-type accuracy, memory-state metrics, token costs
train_sft: honor --base-model and --max-length, drop unused import
CLERK LoRA SFT trainer (TRL SFTTrainer), verified against trl examples/sft_qlora; pushes adapter to the Hub
CLERK SFT dataset builder: timelines -> messages JSONL (CONSOLIDATE + ANSWER mixture), seeded shuffle
CLERK shared prompt definitions (train/eval must never drift)
generator: temporal answer uses slot-held value; gold ADD evicts first when ledger full; budget default 12 for eviction pressure
generator: remove module-global session index; question targets restricted to ledger-supported predicates
common.py: fix facts_of valid_only semantics
CLERK generator: correct single-pass timeline builder, ledger-aware question targets, eviction-safe gold ops
CLERK generator: clean full rewrite of render_session and persona timeline builder
CLERK synthetic evolving-session generator: personas, updates/supersessions/tombstones, gold ledger transitions, seeded
CLERK core: slot schema, edit-op schema, deterministic reducer, serializers
initial commit
