Show HN: Tried some experiments with architecture for Long term memory for LLM

MindCache, an open-source long-term memory system for large language models, launched this week as a Python SDK with an accompanying MCP server, per a Show HN post by its author, Faisal9876. The project is available on GitHub and targets developers building LLM applications that need persistent, structured memory across sessions.

The system organizes memory into four distinct types — user, knowledge, episodic, and decision memories — each with its own lifecycle, role, and token budget within the retrieved context, according to the HN post. Decision memories receive particular treatment: rather than remaining as static, unrelated entries, they can be marked active, superseded, or conditional as they evolve over time. The system uses active decisions as anchors to drive further retrieval via BM25 lexical search, which the author says helps surface relevant context that pure semantic search might miss.

A core architectural choice is what the project calls "smart injection." Instead of assigning new memories to topics purely by embedding similarity, MindCache uses an LLM-guided ingestion step that takes the existing topic structure as context and decides where a new memory belongs and how it relates to prior entries. The author describes this as allowing the hierarchy to grow dynamically rather than accumulating as a flat collection of isolated nodes. Alongside this, MindCache adapts the static RAPTOR-style tree concept into a dynamic hierarchy that updates incrementally as memories arrive, maintaining summaries at each topic level to improve broad queries that might otherwise miss overall context when retrieving individual memories one by one.

On a BEAM evaluation, the author reports MindCache achieved roughly 64% average rubric pass rate compared to approximately 53% for Mem0, with stronger performance specifically on summarization, contradiction resolution, and multi-session reasoning categories, per the HN post. The author also published a Medium writeup providing a broader overview of the project's motivation and design. No third-party validation of the benchmark results is available from the sources provided.

The MCP server component means MindCache can connect directly to any MCP-compatible client, broadening its potential integration surface beyond raw SDK usage. The project is currently at an early experimental stage — the HN post has one point and no comments at time of writing — and the author is explicitly soliciting feedback on the design and implementation.