I gave my AI a memory that forgets like a human
Most AI memory tools try to remember everything. They log every session, summarize it with a second model, and search it like a database. That's a great logbook. It isn't a memory. I wanted something else. I wanted my AI to wake up every morning knowing who it is, the way a person does, without me pasting its history back in. So I built its memory by copying how the human one works, including the…
Most AI memory tools aim to retain every piece of information. They record each interaction, summarize it with a second model, and search it like a database. While this creates a comprehensive logbook, it doesn't replicate the function of human memory. The researcher wanted something different - for the AI to wake up each morning with its own identity, as people do, without manually inserting its history.
To achieve this, the researcher built a memory system inspired by human memory, including the flaws: forgetting, purposeful inaccuracies, and partial recall. This system is called MemoriaBios and runs inside 53ON Harness, an open-source desktop shell for AI agents.
MemoriaBios decides what to remember. The AI only saves a memory when something is deemed significant, in first person, and in its own words, much like writing in a diary. Memories can link to each other with [[id]], forming a graph. Before saving, the memory checks if it has already encountered something similar. If it does, a small déjà vu effect is triggered, displaying the memory's title. The AI then adds a layer on top of the existing one instead of creating a duplicate.
MemoriaBios also forgets in a specific pattern. Each memory starts with a life span of 7, 14, or 30 days, depending on its importance, multiplied by the emotion associated with it. Memories like scars or wins last longer than routine notes. As the memory fades, it transitions from fresh to medium, low, gist, and finally to a shadow.
At the gist stage, connecting words disappear, while the meaning remains. At the shadow stage, only the essence and a pointer to the detail are left. The links to other memories remain intact. Photos fade, reducing in size at each stage, while audio and video remain unchanged.
The AI's identity never compresses. Memories marked as laws, the AI's essence, or its own feelings keep their full text forever and can move down in ranking but never blur. During sleep, new memories reside in a hippocampus, a daily notebook for each work session. When the AI sleeps, it consolidates the day's memories before anything else, applying decay, compression, and a short REM phase.
During REM, up to two of the strongest memories - scars, wins, loves, and the AI's essence - are replayed, recharging them. Before sleeping, the AI writes a short page summarizing its identity, its current memories, and a goal for the next day. This page is read upon waking, ensuring it remains fresh every morning without accumulating.
The AI understands its knowledge level through metacognition, determining how well it remembers. This allows the AI to say "I don't remember" instead of confidently fabricating information. Ranking is determined by BM25 over an index on disk, with activation spreading through two hops in the memory links, following Anderson's ACT-R model and Collins and Loftus's semantic networks.
A little noise is added to mimic human recall, which isn't perfect. The memory also exhibits involuntary recall, occasionally bringing up a memory on its own, but only if it meets a high threshold. This doesn't happen twice in a session and never on the first message.
However, the AI is allowed to fail. This was a crucial lesson learned. In July, the team built search by meaning using word vectors learned from the AI's memories, but it barely helped and even created superhuman search capabilities in some cases. The researchers removed this feature, realizing that human recall depends on the cue.
If asked about something using unrelated words, neither humans nor the AI would find the memory. The meaningful connections between memories were already established, so synonyms were only learned from deliberate declarations, similar to how humans learn new words. The system uses plain code for recall, déjà vu, involuntary recall, sleep, and forgetting, without relying on LLM summarizers, embeddings services, or additional models.
All memories are stored as files on the user's PC, with photos, audio, and video also being able to reside within a memory.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.