kael-bit/engram-rs

Highlights

• Anthropic native API support — Set ENGRAM_LLM_PROVIDER=anthropic to use Claude directly. No proxy needed.
• Per-component LLM routing — Each component (gate, audit, merge, extract, etc.) can have its own provider, model, and API key. Mix cheap models for triage with strong models for judgment.
• Importance-weighted buffer eviction — Buffer overflow now evicts lowest-importance memories first instead of FIFO. Lessons and procedurals are exempt.
• Consolidation dry-run — POST /consolidate?dry_run=true previews what would be promoted, decayed, and demoted without writing anything.
• macOS builds — Pre-built binaries for both x86_64 and aarch64.

What's Changed

Features

• ENGRAM_LLM_PROVIDER env var: openai (default) or anthropic
• Per-component env vars: ENGRAM_{GATE,AUDIT,...}_{URL,KEY,MODEL,PROVIDER}
• Tunable params: ENGRAM_NO_FTS_PENALTY, ENGRAM_HNSW_EF_SEARCH, ENGRAM_TRIAGE_BATCH
• Dry-run response includes would_promote, would_decay, would_demote

Fixes

• Keyword affinity penalty relaxed from 0.7 to 0.85
• Initial importance for lessons/procedurals boosted from 0.5 to 0.75
• Audit interval reduced from 24h to 12h

Full Changelog: https://github.com/kael-bit/engram-rs/blob/main/CHANGELOG.md