kael-bit/engram-rs releases

Breaking Changes

• Rerank removed — rerank field removed from RecallRequest. LLM re-ranking was default-off and unnecessary; FTS + semantic scoring is sufficient. Reduces recall latency.
• Audit → Topic Distillation — old three-powers global audit replaced with focused topic distillation. When a topic grows to 10+ memories, overlapping entries are condensed into fewer, richer ones. More efficient, fewer wasted LLM calls.
• Sandbox removed — no longer needed with the simpler distillation model.

New Features

• Episodic Core block — episodic memories can never be promoted to Core. Enforced at both candidate collection and DB level.
• Web UI split — single 1300-line index.html split into index.html + style.css + app.js. Easier to maintain and read.
• Sort by weight — web UI memory list supports 4 sort modes: Newest, Oldest, Weight ↓, Weight ↑. Uses the same memory_weight() formula as recall scoring.

Improvements

• Centralized prompts — all LLM prompts and JSON tool schemas in src/prompts.rs
• Centralized thresholds — all 60+ magic numbers from 15+ files in src/thresholds.rs
• Epoch-based decay — decay only during active consolidation, not wall-clock time
• Buffer redesign — all new memories enter Buffer; promotion requires consolidation
• Reconcile merge — auto-detects same-topic memories where newer supersedes older
• ~1400 lines removed net across refactoring and feature removal

Breaking Changes

• /recall and /triggers now return clean MemoryResult (id, content, score, layer, tags, kind only — internal fields removed)
• Working memories are never deleted or demoted to Buffer

Unified Scoring

• New memory_weight() function used across all ranking contexts
• Additive formula: (importance + rep_bonus + access_bonus) × kind_boost × layer_boost
• Importance floor changed from 0.3 to 0.0 — memories with high repetition/access stay discoverable even at zero importance
• Recall weights: 0.5 relevance + 0.3 weight + 0.2 recency

Activity-Driven Consolidation

• Consolidation loop skips when no write activity since last run — zero waste during idle periods
• Audit loop follows the same pattern
• Agent frozen = system frozen. No phantom decay while inactive.

Gate Improvements

• Promotion gate now batches all candidates into a single LLM call (was per-item)
• _gate-pending cooldown tag prevents retry storms on LLM errors (2h TTL)
• Triage dedup via _triaged filter

Memory Protection

• Working layer: importance decays to 0 but memory persists — findable via recall and topic tree
• Buffer: still evictable when over capacity
• Core: unchanged (permanent)

Full Changelog: https://github.com/kael-bit/engram-rs/compare/v0.12.2...v0.13.0