Summary

This RFC proposes a new Context Engineering approach grounded in the CIPEA memory framework (Identity / Context / Preference / Experience / Activity) to unify “extract → persist → retrieve → inject” into a single, modular pipeline for LobeChat. It connects a developing & WIP auto distill and auto extract pipeline with a composable Context Engine for message-time injection using Providers and Processors.

Background

Memory is crucial in long-term perspective, ChatGPT, Claude, and many other products implemented their own memory system, ChatGPT Atlas uses browser history as memory, ChatGPT Pulse builds memory from life style, existing infrastructure of Mem0, Zep, and Memobase, memU developed their own memory system to assist agent building to better engineering over contex for agents. Therefore the functionality of memory becomes the essential part for most agents, no matter it's personal assistant, coding-agent or other forms of agent helping human.

Goals

1. Memory Injection: Providers/Processors for consistent context injection and message transforms.
2. Memory extraction: extract memory entries to CIPEA model for a "persist caching layer" for Context Engine.
3. Memory tools: ad-hoc and on-demand memory modification, deep research over user memory, providing a set of wrapped well memory retrieval tools for agent to use.

Design

Key ideas:

• Use the four core memory layers (I/C/P/E/A) to define clear contracts for retrieval and injection.
• Structure context handling with Provider (inject new context) and Processor (transform existing messages) roles.
• We will have a cron/continuously running task to run agent loop (gate → extract → validate → insert) to provide pipeline to auto create memories for user memory entries.
• Provide a practical baseline pipeline with ordering, fallback, and guardrails (token control, dedup, conflict resolution).

Architecture Overview

• Memory Store: A persistent, queryable store for ICPEA entries (any DB + vector index).
• Retrieval Layer: Embedding-based search with re-ranking (similarity × recency × importance).
• Context Engine: Composable pipeline of Providers and Processors for message-time context building.
• Preference Resolver: Scene-based directive resolver with context binding and fallback.
• Control Plane: Configuration, metrics, traces, and guardrails (limits, dedup, conflict resolution).

Core Concepts

• ICPEA Layers: Identity (who), Context (what/where/when), Preference (how), Experience (why), Activity (what happened).
• 5W1H Mapping: Who/What/Where/When/Why/How determines extraction, retrieval, and injection policies.
• Single Source Of Truth: All user-memory-backed context comes through Providers to prevent drift and duplication.
• Deterministic Ordering: Providers run in a stable order (I → C → P → E → A) before final processing/cleanup.

Risks and Mitigations

• Token bloat: truncate-first, per-provider budgets, final cleanup.
• Noise injection: thresholds, dedup, validation/confidence gating.
• Preference conflicts: single directives injection, scene/priority resolution, contradiction tracking.
• Complexity: phased rollout with minimal default set and clear toggles.

Open Questions

• How to evaluate “helpfulness” of injected experiences across domains?
• What governance model for auto-updating preferences from suggestions?
• When and how to surface memory diffs to users for review?