Memory / Persistence
Persistent Memory: Can an AI Remember Over Time?
What survives a restart? F188 and F274 show how Elena AILL tests recall of consolidated information, including questions asked in different words.
What happens to an AI system’s memory when it restarts? Does the information it retained remain available, or is it lost?
Elena AILL has investigated this question through experiments on persistent memory.
In F188, the system faced ten checks after its saved state was reloaded: six recalls using different wording, two links between facts and two recalls without consulting the original document.
It passed all ten checks: 10/10, with a PASS outcome.
A separate experiment, F274, tested semantic recall and links between facts, reaching 10/10 and 2/2 respectively on its own checks.
These results document the retrieval of previously consolidated information under the test conditions. The research continues with how that information is selected and used in answers.
What survives the end of a session?
During a conversation, a system can use information from earlier messages. But what happens when the session ends?
Testing this means restarting the system and checking whether it can still retrieve what it retained.
That is one question explored in Elena AILL’s persistent-memory experiments: whether information remains available and can be retrieved after a restart.
F188: returning to facts without reopening the source
In F188, Elena AILL was tested on information retrieval after a restart. The system had to recover previously stored facts without consulting the original document again.
The outcome is PASS: 6/6 paraphrase-recall checks, 2/2 links between facts and 2/2 source-detached recalls. Those three families make up the 10/10 checks. The test concerns previously consolidated content, not ten new documents learned during the experiment.
F188 · Recall after restart
Six paraphrase checks, two cross-fact checks and two source-detached checks: 10/10 overall in F188. Each bar retains its group’s denominator. The PASS applies to this historical sample, not to every future memory.
Source: F188 · Document Memory Cold Reload ↗Remembering across different wording
Remembering information does not necessarily mean recognising the same sentence every time.
In F188, some questions were therefore reworded while keeping their meaning.
The system passed all six paraphrase checks.
The experiment also included two checks on links between different pieces of information. It passed both.
These findings concern the cases examined in the experiment.
F274: a second observation
A separate experiment, F274, tested information retrieval after the saved state was reloaded. The system passed 10/10 semantic-recall checks and 2/2 checks on links between facts. F274 has its own conditions and results and remains separate from F188. Both experiments document retrieval of previously consolidated information.
What is worth keeping?
Keeping information matters, but a system also needs to know when to use it.
A correct answer depends on retrieving facts relevant to the question. Information can be stored without being useful in every situation.
F188 and F274 answer the questions posed by their historical tests. Unlimited duration, a general absence of forgetting, separation between users and guaranteed deletion need dedicated checks, as does the behaviour of the current service. Neither sample measures those properties.
Recall is a starting point
F188 and F274 show Elena AILL retrieving previously stored information after restarting, under the conditions of their respective tests.
The research continues with a further goal: enabling the system to identify relevant information and use it correctly in its answers.
