Agent memory
Agent memory is what makes an AI agent feel personal. It is the layered store (working, episodic, semantic, core) that records what was said and done across conversations and surfaces, plus the consolidation jobs that promote new memories into long-term storage.
Why it matters
The model is not the moat. Frontier models are weeks apart from each other in capability and the gap is closing. The thing that makes a long-running agent feel like yours, rather than a stateless chat assistant that forgets you between sessions, is memory. The agent recalls what you talked about last week, what you decided, what you do not want to be reminded of, what you are working on, what your kids are named, what your sister is allergic to.
Memory has structure. Most useful agent-memory designs borrow loosely from human cognition: a working memory holding the current conversation, an episodic memory recording what was said and done, a semantic memory holding stable facts, a core memory for the small set of facts you should never forget about the user, and a skills layer recording how the user likes things done. Without these layers an agent is either too forgetful to be useful or so verbose its context window collapses on itself.
The ongoing engineering problem is consolidation: deciding what to keep, what to forget, what to promote to long-term storage, and when. Without consolidation, every conversation either repeats itself or buries old information so deep the agent never finds it again.
How Ajeris handles this
Ajeris memory is layered and consolidated nightly. New memories are written during conversations using memory_save. They are recalled at the start of each turn through embedding-based search plus substring fallback. A nightly dreaming job (3 a.m. UTC, fleet-wide) consolidates important new memories into long-term storage; a daily reflection (11 p.m. UTC, per-user) writes a short summary of the day and may save a skill if the user solved something interesting. Categories include preference, fact, routine, contact, and instruction. The user can save, recall, list, and forget memories at any time by asking.
