most ai chat interfaces are amnesiacs. you open a new tab, type a prompt, have a brilliant 20-minute discussion, and the moment you close the window, all of that context vanishes into the void. when you return the next day, you have to re-explain who you are, what projects you're working on, and how you prefer your code formatted.
when I set out to build nahttyos—my personal ai operating system that lives inside telegram—the core problem wasn't getting the llm to generate smart responses. the real challenge was building a memory system that felt effortless, continuous, and human.
the two types of memory: short-term vs long-term
human memory doesn't store raw video transcripts of every second of our lives. instead, we hold active conversation turns in working memory, while archiving key facts, preferences, and mental models into long-term memory.
in software, short-term memory is simple: it's just the sliding window of the last 10-15 messages sent to the model context API. but long-term memory requires vector embeddings and semantic search.
vector search: finding needles in semantic haystacks
every time a user sends a message or completes a workflow, nahttyos generates a 1536-dimensional vector embedding of that interaction and stores it alongside structured metadata.
when a new query arrives—say, "what was that eth birr stablecoin parameter we discussed last week?"—the system performs a cosine similarity search against previous memories. the top k most relevant historical notes are dynamically injected into the system prompt before calling the inference engine.
why telegram was the perfect interface
web dashboards feel transactional. messaging apps feel persistent. telegram provides an always-on, cross-device message stream with webhooks, push notifications, and background processing.
by placing the ai memory system inside telegram, I can dictate a quick voice note while walking through addis ababa, send a code snippet from my laptop, or receive an automated task execution alert—all within the same continuous chat thread.
lessons learned
- • filter noise early: not every casual "thanks" or "ok" needs to be embedded into long-term memory.
- • decay matters: recency weighting prevents 6-month-old outdated assumptions from overriding current task requirements.
- • keep it quiet: the best ai tools quietly assist you without getting in your way.