When should an agent trust its memory, look for more evidence, or change its mind?
Project Alice is my research on how AI agents remember information and update what they believe. I started with small gridworld experiments and now test related ideas using financial and blockchain transaction data. The goal is to find out whether these choices help agents make better decisions.
Possible explanations
Illustrative example
An agent remembers a transaction, but a new query cannot find it. Which source should it trust?
An independent journal records a later reversal.
Update the current state; retain the original transaction history.
The query used a stale replica. The current journal and receipt agree.
Keep the transaction record and check a different source.
An audit shows the earlier write never committed.
Correct the record and any conclusions that depended on it.
Two agreeing reports use the same cache.
Seek an independent source or leave the transaction unconfirmed.
I compared agents trained only to complete a task with agents also trained to predict their next view. Across three training seeds in this small MiniGrid pilot, prediction training did not help agents remember clues they needed later.
I tested whether agents should change their memory whenever new information contradicts it. In this 38,400-episode study, the main comparison found that selective updates reduced incorrect memory changes, but checking evidence before acting helped agents finish more tasks.
I tested whether learning to predict the next view helps an agent remember a clue it needs later in a gridworld task.
I tested whether checking conflicting reports before updating memory helps agents finish tasks without losing correct information.
A financial reconciliation app that processes transaction files and investigates records that disagree. I use a separate synthetic environment to test memory updates.
A read-only Solana investigator that keeps the evidence behind each finding. I use synthetic transaction histories to test how it handles missing or conflicting information.
On observed-clue trials, task-only models made the correct delayed choice 100% of the time in all three seeds. Models also trained to predict the next view scored 50%, 100%, and 100%; prediction training did not help.
Three training seeds.
Alice made 0.31 incorrect edits per episode but completed 46.6% of episodes, versus 50.8% for action verification.
Main symbolic comparison; 1,920 episodes per controller.
Conventional completion was 95.3%; Alice completion was 57.8%. Alice made no harmful revisions in this comparison.
32 families; 10,752 records across four budgets; metrics at budget 10; controls and non-identifiable cases excluded.
Alice completed 32.8% of cases, compared with 33.3% for conventional reconciliation. Its weighted harmful-revision rate was 62.5%: some initially correct memories were changed.
Separate 32-family run; 10,752 records across four budgets. Budget-10 results exclude controls and cases without enough evidence to identify the cause; conditions and families receive equal weight.
In the primary closed-loop comparison, Alice and the deterministic policy both completed 62.5%. Alice used 2.25 queries on average versus 1.75.
32 episodes per policy in this comparison; 704 across all three studies. Two synthetic trace seeds; no held-out test trace.
How should an agent check a claim when several sources share the same failure?
When should an agent gather more evidence, revise a belief, or abstain?
Do selective memory updates help in new environments or when an agent’s own capabilities change?