Engram

• Brain-inspired, portable context database for AI agents that stores knowledge as structured “bullets” in a concept graph rather than raw text ( softmaxdata/engram).
• MIT-licensed, maintained by softmaxdata
  • 19 stars

Any AI system can access and build upon shared context regardless of which LLM or framework powers it ( softmaxdata/engram).

• Cross-model portability — context transfers between Claude, GPT, Gemini, and other LLMs unchanged
• Atomic knowledge units — each bullet is individually tracked with usage statistics
• Reconsolidation — helpful bullets strengthen over time, unhelpful ones fade (memory as a living, biologically-inspired substrate)

• Core memory slot — always-loaded summary (≤512 tokens) preserved across context compaction
• Worked-example retrieval — attaches prior similar inputs together with their successful outputs near new queries
• Diversity ranking — Maximal Marginal Relevance prevents token budgets from filling with near-duplicates

• Separates read (materialization) from write (commit) so retrieval never blocks on ingestion
• Server-side “Reflector” LLM processes all raw input canonically, so bullets are consistent across clients
• Three storage tiers: active, archived, purged

• Reflector — canonical LLM for knowledge extraction
• Curator — decides what gets stored via deduplication and validity checks
• Delta engine — applies atomic mutations, never full rewrites
• Activity ledger — permanently retains raw text for future re-extraction