ED Resolve

A developer shared their project ED Resolve on Reddit, with the core problem being enabling an AI system to remember past contexts and draw on experience when making decisions, while also proactively discarding those memories at the appropriate time.

ED Resolve uses a component called Hindsight to handle the responsibility of memory storage.

ED Resolve's decision chain is: current context → Hindsight memory retrieval → judgment of whether the memory is applicable → generation of possible actions → constraint checking → simulation and inference → final decision.

After retrieving historical experience, ED Resolve first performs a relevance assessment; if the difference between the current context and the memory context is too large, the system chooses not to rely on that memory.

ED Resolve's design repositions memory from 'instruction' to 'evidence,' where evidence is merely an input for decision-making reference and must pass constraint checking and simulation-based verification.

ED Resolve adds a layer of meta-evaluation on top of RAG for retrieved results, rather than directly injecting retrieved content into the prompt for the model to execute.

ED Resolve introduces a simulation and inference step that runs candidate actions through inference to observe which one genuinely makes sense under current conditions.

ED Resolve's current sharing remains at the level of architectural description and dashboard display, lacking quantitative evidence regarding applicability criteria, simulation inference cost accuracy, and comparisons of actual effectiveness.

Decoupling memory retrieval, applicability judgment, constraint checking, and simulation inference into independent steps so that each step is auditable and intervenable is a referenceable idea offered by ED Resolve.