longbrain Files

Highlights

• KG-lite: triple-based fact supersession + entity graph, with strict/boost/global project recall scope
• Memory quality: contradiction detector for conflicting facts, open/done status for task-type facts, last_seen refresh so recalled facts resist decay, tightened fact-extraction criteria against pollution
• Consolidation reliability: fixed data-loss/scope-leak gaps, corrected retry idempotency and backlog accounting, added a turn_id idempotency key, fixed pair-retry duplication and the oversized-turn cap, deterministic repeat-occurrence disambiguation, complete VN diacritic set
• UI: Memory Atlas redesign with semantic graph layers and a health dashboard, aligned with the Longbrain theme, plus stability/perf fixes
• Codex adapter: official lifecycle hooks (auto recall, recording, session-start catch-up) and an AGENTS.md memory-priority patch

Full commit list

• feat(ui): align Memory Atlas with Longbrain theme
• fix(memory): wire turn_id through every remaining /memory/append caller
• fix(memory): add turn_id idempotency key, closing the anchor heuristic's gap
• fix(memory): make repeat-occurrence disambiguation deterministic, not time-based
• fix(memory): stop retry from reopening consolidated turns, complete VN diacritic set
• fix(recall): match context_block section headers to the query's language
• fix(memory): fix pair-retry duplication, summary regression, and oversized-turn cap
• fix(memory): correct retry idempotency, force-turns trigger, and backlog accounting
• fix(memory): close data-loss and scope-leak gaps in consolidation and recall
• docs(ui): add Memory Atlas screenshot to README
• feat(codex): patch global AGENTS.md with a memory-priority instruction
• fix(memory): tighten fact-extraction criteria against pollution
• feat(kg-lite): strict/boost/global project recall scope
• fix(ui): stability and perf fixes in the memory atlas
• feat(ui): redesign the memory atlas
• fix(kg-lite): close 3 supersession reliability gaps from code review
• fix(hooks): skip Codex guardian/subagent rollouts in turn-ended sync
• test(evals): add genuine multi-hop gauge case for KG-lite's entity graph
• feat(ui): graph semantic layers — topic sub-clustering
• feat(ui): memory health dashboard
• feat(memory): contradiction detector — flag conflicting facts
• feat(memory): open/done status for task-type facts
• feat(memory): last_seen refresh so recalled facts resist decay
• feat(adapters): Codex official lifecycle hooks — auto recall, recording, session-start catch-up
• fix(evals): pass triple fields through recall_eval's seeding harness
• feat(memory): triple-based fact supersession + entity graph (KG-lite)

Full diff: https://github.com/ngocthanh06/longbrain/compare/v0.1.0...v0.2.0