| # TMCRA Training Notes |
|
|
| This document summarizes the training direction behind the graph scorer package included in this release. |
|
|
| ## Training Goal |
|
|
| TMCRA trains graph-scoring components for long-memory retrieval. The goal is to help the runtime decide which memory nodes and graph paths should be surfaced to an answer model for a given user query. |
|
|
| The trained model is not intended to replace the answer LLM. It is responsible for memory selection: |
|
|
| - identify relevant memory nodes |
| - score graph paths between related memory events |
| - preserve useful cross-turn and cross-session links |
| - reduce noisy or stale evidence before answer generation |
|
|
| ## Model Components |
|
|
| The released model directory contains two main runtime scorers: |
|
|
| - `node_scorer.pt`: scores candidate memory nodes. |
| - `path_scorer.pt`: scores graph paths and tunnel links between memory nodes. |
|
|
| The training output also includes: |
|
|
| - best checkpoints |
| - last checkpoints |
| - epoch and step checkpoints |
| - training summary and logs |
| - export manifest |
|
|
| ## Training Data Direction |
|
|
| Training data is built around dialogue-memory behavior rather than isolated QA pairs. Samples are designed to teach the model how memory should connect across turns and sessions. |
|
|
| The major training directions include: |
|
|
| - direct user facts |
| - assistant-provided details |
| - preference/profile signals |
| - temporal state changes |
| - old-value vs current-value selection |
| - cross-session event links |
| - multi-evidence aggregation |
| - evidence-positive vs noise/negative memory separation |
| - unit-to-unit coverage for count/sum/compare tasks |
|
|
| ## Graph Memory Supervision |
|
|
| Each training example is converted into graph-oriented supervision. Instead of only asking whether a text chunk is relevant, TMCRA trains over: |
|
|
| - memory node relevance |
| - event-unit relevance |
| - path usefulness |
| - tunnel/link usefulness |
| - evidence role |
| - currentness and temporal state |
| - whether a candidate should be injected into answer context |
|
|
| This allows the runtime to learn memory structure, not only lexical similarity. |
|
|
| ## Writer and Scorer Separation |
|
|
| TMCRA separates memory writing from graph scoring. |
|
|
| The writer extracts candidate memory records from dialogue. The graph model then learns how those records should be selected and connected during retrieval. |
|
|
| This separation is important because a long-memory system needs two different abilities: |
|
|
| - write useful memory units from conversation |
| - retrieve and connect the right units later under noise |
|
|
| ## Training Output Included |
|
|
| The packaged model output is located at: |
|
|
| ```text |
| models/action_frame_tunnel_graph548_tunnel_fusion_train_20260524_042557/ |
| ``` |
|
|
| Runtime files: |
|
|
| ```text |
| node_scorer.pt |
| path_scorer.pt |
| export_manifest.json |
| ``` |
|
|
| Full training trace: |
|
|
| ```text |
| checkpoints/ |
| node_scorer_best.pt |
| path_scorer_best.pt |
| node_scorer_last.pt |
| path_scorer_last.pt |
| train_summary.json |
| train.log |
| training_issues.jsonl |
| ``` |
|
|
| ## Current Training Lessons |
|
|
| The current baseline shows that TMCRA has strong single-session fact recall and assistant-detail recall. It also has working temporal and preference layers. |
|
|
| The main remaining training targets are: |
|
|
| - stronger multi-session aggregation |
| - better unit coverage for count/sum/compare questions |
| - deeper temporal graph planning |
| - query-graph to memory-graph matching |
| - more stable preference abstraction under indirect user requests |
|
|
| These directions are the next step for improving TMCRA from a working long-memory runtime into a stronger general agent-memory layer. |
|
|
