I’ve been working on Engram, a persistent context layer for AI agents and LLM sessions.

The inspiration came from a simple problem: when you work with LLMs for a long time, context collapses. Conversation history gets too long, summaries drift, and models forget or mix up earlier decisions. It becomes even worse when multiple agents or different models are involved.

Engram is an attempt to treat AI memory more like how human memory works: * knowledge is stored incrementally * related concepts reinforce each other * context is reconstructed when needed rather than replaying the entire past

Instead of storing conversations as raw text or chunks, Engram commits information into a persistent context structure and then materializes the relevant pieces when a model needs them.

Some other capabilities built so far: * multi-agent updates to the same context * evolving knowledge instead of append-only memory * context reconstruction across long sessions * working across models (e.g. start in ChatGPT, continue in Claude)

Repo: https://github.com/softmaxdata/engram

Would love feedback from people building agents or long-running AI workflows. Happy to answer questions or run experiments people suggest.