Feature Proposal: Context-Aware Intent Memory Chain

[mirror-in-peace]

Currently, intent-fluid treats each human intent as a relatively isolated execution block. While effective for simple tasks, this "stateless" approach limits the system's ability to handle complex, multi-step workflows where subsequent intents rely on previous context or user preferences.

I propose implementing an Intent Memory Chain to enhance the autonomous execution flow:

1. Short-term Session Memory (Context Buffer)

• The Problem: If a user says "Search for Nvidia's stock price" followed by "Now analyze it," the second intent lacks the "Nvidia" context.
• Proposal: Maintain a sliding window of the last $N$ intent-skill execution results to provide immediate context for the next LLM routing decision.

2. Long-term Preference Memory (RAG-based)

• The Problem: Recurring intents (e.g., specific output formats or preferred tools) are re-processed from scratch every time.
• Proposal: Use a lightweight vector store to persist successful execution patterns and user-specific constraints.

Why this matters:

• Reduces Ambiguity: Better intent resolution by filling gaps with historical data.
• Smoother "Fluid" Experience: Minimizes the need for redundant user prompts.
• Higher Execution Confidence: Provides the agent with a "mental model" of the ongoing task.

Is this something aligned with the current roadmap? I’d be happy to discuss the implementation details or contribute a PR if there's interest in this direction.

[atalia]

Thank you for putting together this detailed and well-thought-out proposal!

You hit the nail on the head regarding the limitations of the current "stateless" approach. Transitioning to a context-aware system is definitely something we want to explore, as it aligns perfectly with our goal of making the execution flow more intuitive and "fluid" for complex tasks.

A few quick thoughts on your proposed directions:

Short-term Session Memory: A sliding window approach for the context buffer sounds like a solid, pragmatic first step. It should immediately resolve the core ambiguity issues you mentioned.

Long-term Preference Memory: Using RAG to persist successful patterns is very interesting. One thing we will need to consider during implementation is keeping the dependencies as lightweight as possible when selecting a vector store.

We would absolutely love your help with this. Before jumping straight into the code, would you be open to creating a brief design document or opening a Draft PR? It would be great to see a high-level overview of the proposed architecture (especially how the memory modules will interface with the current LLM routing logic) so we can align on the technical details.

Thanks again for driving this forward. Let me know how you'd like to proceed!

[carbonshow]

Thank you for putting together this detailed and well-thought-out proposal!

You hit the nail on the head regarding the limitations of the current "stateless" approach. Transitioning to a context-aware system is definitely something we want to explore, as it aligns perfectly with our goal of making the execution flow more intuitive and "fluid" for complex tasks.

A few quick thoughts on your proposed directions:

Short-term Session Memory: A sliding window approach for the context buffer sounds like a solid, pragmatic first step. It should immediately resolve the core ambiguity issues you mentioned.

Long-term Preference Memory: Using RAG to persist successful patterns is very interesting. One thing we will need to consider during implementation is keeping the dependencies as lightweight as possible when selecting a vector store.

We would absolutely love your help with this. Before jumping straight into the code, would you be open to creating a brief design document or opening a Draft PR? It would be great to see a high-level overview of the proposed architecture (especially how the memory modules will interface with the current LLM routing logic) so we can align on the technical details.

Thanks again for driving this forward. Let me know how you'd like to proceed!

thx @atalia , adding long/short-term memory support could significantly enhance contextual understanding and improve the accuracy of problem-solving. If you're interested, feel free to help tackle this issue.