[ Feature Request ]: Event-Driven Memory (Selective Storage of Meaningful Messages)

[sobowalebukola]

Currently, the memory layer stores data more uniformly, which can lead to noisy context and unnecessary persistence of low-value interactions. This feature proposes an event-driven memory pipeline that selectively persists only “meaningful” user events (e.g., goals, preferences, commitments, corrections, achievements) based on heuristic or model-based signals.

Motivation

• Reduces memory bloat and storage overhead

• Keeps user memory more relevant, concise, and high-value

• Improves personalization without degrading context with trivial chats

• Aligns with real-world memory patterns (event-based rather than every utterance)

Proposed Behaviour

The system should:

• Process each message through an event classifier

• Decide whether it should be saved, discarded, or summarized

• Optionally create state updates (overwrite old info instead of accumulating)

Examples of “meaningful events”:

User expresses long-term preference (“I prefer dark mode”)

User sets a goal (“Remind me to study Rust this month”)

User corrects facts (“My role is actually backend engineer”)

User shares achievements or milestones

User updates personal profile info

Technical Notes / Implementation Ideas

• Implement a lightweight classifier (rules + LLM or embedding similarity)

• Add memory entry metadata (weight, type, expiry, source)

• Store only high-weight events

• Consider TTL for certain memory types

• Add hooks so devs can customize “event” categories

Potential Challenges

• Defining the signal for “meaningful”

• Updating existing stored knowledge without duplicating

• Balancing privacy with persistence

• Preventing accidental storage of highly personal or sensitive info

Acceptance Criteria

• Low-value messages are not stored by default

• Significant user info persists reliably across sessions

• Memory size growth slows or stabilises

Additional Context

This feature aligns with feedback around improving personalization while avoiding “creepy” over-storage, as well as better mirroring human memory patterns.