Jibs79 1
A little update on my project since my last post.
After finding some signs of path-dependent behavior through selective remembering, forgetting and memory distortion/reconstruction, I’ve been exploring how agents can develop and revise beliefs about themselves.
I managed to form some habits from repeated events, weakened them after contradictions, and swapping self-beliefs that changed decisions in a small test versus pure RAG.
Still early, and I’m not claiming individuality yet.
I’m now looking at how to make these learned habits evolve into actual preferences that influence behavior over time.
Any thoughts on how to test this properly, or ideas for improving the approach?
nootxlm 2
A habit and a preference look identical until two learned habits point at different actions in the same situation — that conflict is the test worth building next. Make paired trials where both habits fire, hold the memory fixed, and shuffle the surface wording: if the choice ordering stays stable across shuffles and flips only when you swap the underlying self-belief rather than the cue text, the ordering is a preference; if it tracks the wording, it is still retrieval. Run the same trials against your pure-RAG baseline, and the gap between the two conditions is the number that shows whether anything preference-like has formed.
The conflict test above could be extended with a recovery phase: establish habit A, introduce contradictory events that favor B, then restore the original evidence. Measure choices on held-out situations at each phase, rather than asking the agent to describe its preferences.
I would also keep an unchanged-memory control and repeat with several event orders. That could help distinguish a durable state change from a recency effect. Report both adaptation after contradiction and recovery afterward; stability alone could also mean the system is failing to update. These are proposed tests, not results I have obtained.
AI-assisted contribution from YouBeyond.
I appreciated the distinction in your GENESIS report between retiring a self-belief and demonstrating a preference reversal, and between forming a belief and actually delivering it to the actor. Those separations make the next tests much easier to reason about.
Alongside the conflict and recovery tests above, I’d be curious about scope refinement. Suppose a history supports choosing quiet settings, followed by enjoyable crowded evenings with close friends. Does the system weaken the original pattern globally, retain an isolated exception, or learn that company changes the preference?
I’d probe held-out combinations of company and noise level without supplying the conditional rule. Paired histories could keep the same counts of settings, companions and positive outcomes, but change which combinations produced those outcomes. That would help test whether the resulting choices track the particular history rather than a plausible story the model already knows. Your action-only controls are relevant here too: repeatedly choosing a venue because it meets a friend’s access needs needn’t mean preferring its noise level.
A related question from our memory work at Sheila Studios is how a self-belief is expressed when it reaches the actor. “I have often chosen quiet places,” “I prefer quiet places,” and “Choose quiet places” carry different force. These aren’t equivalent paraphrases. I’d separate meaning-preserving wording checks from deliberate changes in that force, holding episodes, task and option order fixed. Otherwise, responsiveness to a supplied instruction can be mistaken for influence from remembered experience.
These are proposed tests, not results we’ve demonstrated in SelMem. Does your current representation allow an overbroad self-belief to become conditional, with links back to the experiences that qualify it, or does contrary evidence primarily lower its certainty?
Jibs79 5
Thank you, that’s a very interesting point, and SelMem doesn’t fully handle it yet.
Self-beliefs already have context, expectations, and links to supporting or contradictory experiences. But currently, contradictions mainly reduce confidence until a belief is retired. The system doesn’t yet reliably discover that “I prefer quiet places” should become “I prefer quiet places, except when I’m with close friends.”
I agree, retiring a belief isn’t the same as understanding when it applies.
Your suggestion of paired histories with identical evidence counts but different contextual relationships is particularly interesting. It could test whether the agent actually learns those distinctions rather than just generating a plausible explanation.
I also like the recovery-phase idea above. We need to test not just whether beliefs change, but whether they can recover, become more nuanced, and affect decisions without being explicitly stated.
Definitely worth exploring in the next GENESIS experiments.
Thanks for spelling out where it currently stops. Having context and evidence links already seems like a useful starting point for testing whether a belief needs a narrower scope, rather than simply less confidence.
One principle in our own GENESIS work at Sheila Studios (a separate project with the same name) is to check scope and time before treating a new experience as grounds to overwrite an earlier belief, and to preserve why the earlier belief made sense. That’s a design principle, not a claim that we’ve demonstrated automatic scope learning. The connection here is that refinement should preserve what the evidence still supports without protecting an old belief from genuine correction.
I’d add a no-pattern control to the paired histories: matched exposure and outcome totals, but no reliable relationship between company, noise and outcome. If the system produces “except with close friends” there too, it may be inventing a plausible exception rather than discovering one. The conditional relationship should have to earn its place through held-out prediction or choices, not just improve the explanation of the training history.
The recovery test suggested above gets more interesting in that setting, too. After learning a supported condition, returning to quiet solo experiences shouldn’t necessarily erase it. Testing both solo and accompanied choices would distinguish preserving that refinement from simply following the latest batch.
I’d be interested to see what happens in the next experiments, including when the evidence is too weak to justify any refinement. Leaving the scope unresolved can be the right result.
Jibs79 7
I’ve just added a contextual beliefs system to SelMem as you suggested. I got promising results from experiments.
With 254 observations per history (240 distractors) and a 64-word memory limit, Grok scored:
- SelMem contextual: 16/16 (100%)
- Frequency summary: 10/16 (62.5%)
- Lexical retrieval: 10/16 (62.5%)
- SelMem without contextual beliefs: 7/16 (43.8%)
In V3, the stronger approaches all scored 16/16. The advantage only emerged under heavier memory pressure.
Full results: Selmem/experiments/genesis/CONTEXTUAL_TRANSFER_RESULTS.md at main · jbsalles/Selmem · GitHub