Interactive Memory Learning for Long-Term Conversations

简体中文

L1 observation-safe Planner/Trigger delayed-reward state machine; no paper-scale dialogue training.

Paper information

Field	Details
Paper	Interactive Memory Learning for Long-Term Conversations
Institution	Harbin Institute of Technology, Shenzhen / Pengcheng Laboratory (first author).
First public release	2026-09-15 (arXiv v1)
Author code	No public author repository found as of 2026-09-16.
Local adapter or method key	interactive-memory
Local reproduction code	src/auto_research/agent_research/latest_20260916_followup.py

Paper method and evidence

The Planner selects high-value information to write, the Trigger decides when to retrieve it, and delayed cross-session reward trains the two policies together.

flowchart LR
  P[Planner chooses what to store] --> M[Memory state]
  M --> T[Trigger chooses when to retrieve]
  T --> R[Delayed cross-session reward]
  R --> P

Original paper figure

Figure 1 shows the planner, trigger, and cross-session interaction. From the original paper; copyright remains with the authors.

Core equation

The local reference exposes gate and state changes for fixture checks.

Reported results

Paper accuracy, training losses, and benchmark results are not directly comparable to the local diagnostic.

Local reproduction

Three-seed mechanism results are diagnostic_only=true. Fixed-budget fixtures verify the defining operator and auditable intermediate states; they are not a benchmark, production result, or capability-ranking entry.