stabilize passive memory bench traces

track passive memory followups

make memory folgezettel-native

track memory bench regression

complete planner-grade memory evidence assembly

implement session-first memory retrieval profile

remove hardcoded memory query cues

organize memory architecture docs

cache bench embeddings locally

add planner-grade evidence assembly

improve longmemeval qa bench ergonomics

feat: move global embed cache path to benchmarks/runs/

feat: use persistent data dir for global embed cache

add tree-sitter code intelligence tools

fmt

fix finish reason thingy

fix: cap dense embedding texts to fit safely under 1024 tokens

docs: update beads task completion log

feat: implement principled scoring and adaptive neighbor expansion

feat: integrate BGE-M3 sparse retrieval integration path

refactor: introduce EvidenceAtom, replacing RetrievedRef candidates

bench: wire --official-eval-cmd into longmem subcommand

- relax positional-only arity check to also accept --official-eval-cmd flag - fix positional extraction to skip flag key+value pairs (not just keys) - call evaluator after the loop, read official_eval.json, print accuracy - default LLM URL was localhost:8001 (dead); pass proxy URL positionally smoke result: 0.9333 accuracy on subset30 (28/30), 1.6s avg hot query

bench: add evaluate_qa_local.py local judge for LongMemEval QA eval

reimplementation of official evaluate_qa.py that routes through the local antigravity proxy (127.0.0.1:8000/v1) instead of requiring OPENAI_API_KEY. same prompt templates and positional interface. - scripts/evaluate_qa_local.py: standalone script, nix-shell -p python3Packages.openai for deps, stream=True for SSE proxy compat - benchmarks/README.md: 'Official QA Evaluation (local judge)' section with --official-eval-cmd invocation for smoke (subset30) and full runs closes klbr-1yn (local judge available, no external dep needed) unblocks klbr-u3t

read memory provenance from canonical ref edges

add continuous-loop memory benchmark

improve tool-required router separation

gate graph-only evidence packets

refresh router calibration artifacts

calibrate router tool recall

implement evidence packet constraints

update beads issues export for klbr-f0p

remove tools lane and refactor router to 2-class setup

docs: compile comprehensive benchmark issues and gaps report

Close memory architecture epic

Validate locomo adapter smoke

Add source-grounded memory note writer

Write structured episode event cards

Clarify memory packet reader prompt

Update memory architecture docs

Rerank memory evidence packets

Order turn-window evidence chronologically

Add lane route tracing

Record qa eval follow-up

docs: document klbr.kdl and bench.kdl config usage/secrets and close klbr-d67

docs: update AGENTS.md and close klbr-djb

Document memory regression preflight

Add packet evidence metrics

Route passive recall through ref packets

Share memory evidence planner

Constrain lexical memory retrieval to FTS