Agent memory
Persistent state an agent carries across sessions: facts about its user, corrections received, patterns that worked. Without it every run starts from zero; with it the tenth run outperforms the first.
The cost of starting from zero
An agent without memory re-derives everything it needs on every run: your naming conventions, your preferences, the correction you gave it yesterday. That is expensive in tokens and worse in quality, because the correction never sticks. Memory is what makes an agent improve with use rather than merely repeat.
Not all memories are the same
It helps to type them. Facts about the user, preferences about how work should be done, corrections received, patterns that worked, background context on a project — each is retrieved in different circumstances and deserves different weight. A flat pile of undifferentiated notes retrieves badly and grows without bound.
Retrieval is the hard half
Writing memory is easy; deciding what to surface for the task at hand is the engineering. Load too much and the context window fills with noise that degrades the very reasoning it was meant to improve. Selective, relevance-scored retrieval beats a large memory every time.
In AGNT
AGNT stores typed memories — facts, preferences, corrections, insights — locally, and retrieves them when relevant.
Agent memory — common questions
How is agent memory different from RAG?
RAG retrieves from a document corpus to ground an answer in external material. Memory is the agent’s own accumulated state about you and its past work — what it learned rather than what it read.
Where is AGNT’s agent memory stored?
On your machine, alongside the rest of the runtime. Memories are typed, searchable and yours — they are not a hosted profile held by a vendor.
Can I correct something an agent remembered wrongly?
Yes. Corrections are a first-class memory type in AGNT precisely so they outrank the mistaken belief they replace.