how to build an agent that never forgets

Rohit, in a recent thread, describes how a failed interview pushed him to rethink agent memory, and the write-up is a practical wake-up call for anyone building assistants. You can read the original thread here: https://x.com/rohit4verse/status/2012925228159295810.

He starts with a simple truth many of us have felt, that storing conversation history or embeddings alone looks like memory, but it breaks fast when scale, time, and contradictions arrive. The post walks through a mental shift, treating memory as infrastructure not a feature. That alone is worth pausing over.

The piece outlines a clear stack. For short-term continuity Rohit uses checkpointing (snapshots of agent state), which gives determinism, recoverability, and debuggability. For long-term memory he proposes two architectures: File-Based Memory (a three-layer system of resources, items, and evolving category summaries) and Context-Graph Memory (a hybrid of vector discovery plus a knowledge graph for precise facts). He explains how active memorization rewrites summaries when facts change, so the system knows someone switched jobs, rather than inventing a blended hallucination.

He also shares operational lessons that are often overlooked, like memory decay, nightly consolidation, weekly summarization, and monthly re-indexing. These maintenance jobs keep memory useful instead of noisy. The retrieval strategy he describes uses a synthesized query, relevance scoring, and time-decay, so recent, slightly less similar memories often win over ancient perfect matches.

If you’ve ever watched an assistant confidently get a user’s history wrong, this thread reads like the repair manual. It’s practical, a bit opinionated, and full of patterns you can implement. Expect better companions when memory is treated like an operating system, with RAM for the now and a structured hard drive for the long haul. The future he sketches is optimistic, because structure, not just scale, will make agents that really remember.