Engram: The Memory Revolution LLMs Have Been Waiting For

Long-term memory is the Holy Grail for large language models (LLMs), but current solutions fail balancing cost, speed, and accuracy. Engram emerges as a breakthrough with its dual-process memory engine that dramatically enhances memory capabilities while slashing token consumption.

Engram's Innovative Approach

Traditional LLMs suffer from forgetfulness, a critical flaw when using them across sessions. The conventional fix, replaying the entire history, proves impractical due to its expense and inefficiency. Enter Engram, which discards this clunky method for a smarter, open-source design. It employs a bi-temporal data model that decouples memory storage from LLM processing, offering a substantial edge in usability.

Engram saves memory in two paths, speeding up the process and improving accuracy. Its fast write path records episodes without LLMs, while an asynchronous path crafts a bi-temporal knowledge graph, keeping all facts intact and traceable. Engram doesn't delete contradictory facts. it marks them as invalid while preserving their history and context.

Benchmarking the Uncontestable Gains

Performance speaks volumes. In the LongMemEval_S test, Engram's lean setup scored an impressive 83.6%, outperforming the full-context approach by over 10 points. This wasn't just a small token saving, Engram used only 9.6k tokens compared to the traditional 79k, an 8-fold decrease. This is a quantum leap in efficiency that demands attention.

Why's this important? Because memory benchmarks often get muddled in inconsistent, non-reproducible results. Engram stands out by providing a transparent, in-repo evaluation harness and raw logs. It sets a new standard for reliability and transparency in LLM memory evaluation.

The Future is Now

Engram's hybrid read path, a blend of dense, lexical, graph, and contextual signals, achieved what seemed impossible. It offers a compact and provenance-tagged context, outperforming traditional memory practices without sacrificing detail. By fusing facts with retrieved data, Engram ensures nothing slips through the cracks.

So, if memory is the missing piece in the LLM puzzle, isn't Engram the answer we've been searching for? It's time to ask why every LLM isn't using such a system already. Engram proves that AI memory, you really can have your cake and eat it too.