Memory is strange.

AI memory systems have an accuracy crisis. Recent benchmarks show answer accuracies below 56%, with hallucination and omission rates remaining high. Most systems use LLM extraction on every message, compounding errors until the original truth is lost.

Engram takes a different approach: memory you can trust.

What it does

Engram is a memory system for AI applications that preserves ground truth and tracks confidence:

  • Store first, derive later — Raw conversations stored verbatim. LLM extraction happens in background where errors can be caught.
  • Track confidence — Every fact carries provenance: verbatim (highest), extracted (high), inferred (variable).
  • Verify on retrieval — Applications filter by confidence. High-stakes queries use only trusted facts.
  • Enable recovery — Derived facts trace to sources. Errors can be corrected by re-deriving.

Memory types

Six memory types, each with different confidence and decay characteristics:

  • Working — Current conversation context, volatile and in-memory
  • Episodic — Immutable ground truth, verbatim storage
  • Factual — Pattern-extracted facts (emails, dates, names)
  • Semantic — LLM-inferred knowledge with variable confidence
  • Procedural — Behavioral patterns and preferences
  • Inhibitory — What is explicitly NOT true

The philosophy

Ground truth is sacred. Every derived memory points back to source episodes. If extraction makes a mistake, re-derive from the original. Forgetting is a feature—memories decay over time, keeping the store relevant.

Status

Engram is in pre-alpha. Built on Pydantic AI with durable execution (DBOS/Temporal) and Qdrant for vector storage.

View on GitHub →