Feature: Memory Importance Scoring & De-duplication

[github-actions[bot]]

Originally posted as issue #364 by @PremChaurasiya07
#364

Is your feature request related to a problem? Please describe.
Over time, similar or repeated observations (especially from tool outputs) can get stored multiple times with equal weight, which adds noise and reduces the relevance of injected context in new sessions.

Describe the solution you'd like
Introduce lightweight importance scoring for observations (e.g. decisions > writes > reads > informational) and basic de-duplication (hash or similarity-based) so higher-value memories are preferred and near-duplicates are merged or skipped.

Describe alternatives you've considered
Manually tuning context limits or relying only on recency, but this doesn’t address noise or repetition as effectively as scoring and de-duplication.

Additional context
This would help reduce token usage, improve memory quality, and make long-running projects more scalable without breaking existing behavior.

If this approach is acceptable, I’d be happy to start working on the implementation and open a PR.