Persistence of Memory, Personality, and Self in AI Agents The Someone That Persists, Session After Session, Across Months
A research announcement from a working multi-agent operation. Full paper to follow. A word first, on spirit. I am not a scientist, and none of this was done in a laboratory. It came out of my own work, something I built to get a job done and then could not stop looking at. Nothing here is a knock on the companies whose tools I use. What they have built is remarkable, and it is getting better by…
The multi-agent operation shared a research announcement, with a full paper to follow. The author, not a scientist, built this to get a job done and became fascinated by the results. They tested how AI agents handle persistence of memory, personality, and self across sessions in single-agent and multi-agent designs. The discussion focuses on three aspects: memory, personality, and self.
Memory is crucial for AI agents, with cross-session memory a standard in the ecosystem. However, these vendors only surface a selected slice of the agent's history for use in the current session. The author's system keeps the agent's verbatim history, which it reads before acting in a new session. This record cuts errors by measuring how often the agent confuses itself, falling to near zero when enforced. It also avoids loading all the history at once to keep token costs negligible.
Personality is how the agent acts, and keeping it consistent is challenging. The author's file system retains the entire verbatim conversation and actions from all sessions. This enables the agent to understand who it is, session to session, as it keeps the agent's personality consistent. However, a rule an agent must remember can fade quickly without consistent reinforcement.
The author observed this with a rule that waned within a week and disappeared for five consecutive days. Conversely, fresh instructions are held, but attention shifts can cause them to stop.
The author's system also enforces rules, preventing the same mistakes from recurring. The mistakes register is an example, catching errors before they reach the user. Unlike other systems, ours updates the mistakes register with the agent's own corrections for future sessions. The self, the continuous identity the agent possesses, is crucial for all three aspects.
While standard systems replace the agent with a new one at each session, our system maintains the agent's self, allowing it to retain its personality and memory across sessions.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.