Upload README.md with huggingface_hub

README.md

@@ -10,46 +10,41 @@ tags:
   - agent
 ---
 
-# Memory-R2 7B — Memory Manager + Answer Agent
+# Memory-R2 7B — Memory Manager
 
-Checkpoints from [Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents](https://arxiv.org/abs/2605.21768) (arXiv:2605.21768).
+The trained **memory-management policy** from [Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents](https://arxiv.org/abs/2605.21768) (arXiv:2605.21768). This is the paper's main contribution and deployed "champion" (`32sess_champion_v2`, LoGo-GRPO curriculum, global step 5).
 
-This repo contains **two separate Qwen2.5-7B-Instruct checkpoints** that play two different, non-overlapping roles in the pipeline. They are not interchangeable and neither one does the other's job:
+It is a Qwen2.5-7B-Instruct model fine-tuned with **LoGo-GRPO** (turn-level + token-level credit assignment) via a curriculum of 8 → 16 → 32-session rollouts on the LoCoMo long-horizon dialogue dataset. Given a running conversation, it decides what to INSERT / UPDATE / DELETE in an external memory store.
 
-| Subfolder | Role | Does it manage memory? | Does it answer questions? |
-| --- | --- | :---: | :---: |
-| **`memory-manager/`** | Stage 1 — reads the running conversation and decides what to INSERT / UPDATE / DELETE in the external memory store. | ✅ Yes — this is its only job | ❌ No — it never sees or answers the held-out QA questions |
-| **`answer-agent/`** | Stage 2 — given a question and the memory store `memory-manager/` produced, generates the final answer. | ❌ No — it never touches the memory store's write operations | ✅ Yes — this is its only job |
-
-**`memory-manager/`** is `32sess_champion_v2` — the LoGo-GRPO-trained (turn-level + token-level credit assignment, curriculum 8→16→32 sessions on LoCoMo) memory-management policy, and is the paper's main contribution / "the champion" checkpoint. It is the piece you need regardless of which answer agent you pair it with.
-
-**`answer-agent/`** (`sft_cont_step55`, SFT+RL-trained) is an *optional, swappable* stage-2 model — the memory manager was evaluated in the paper against several different answer agents, and this is simply the best-performing one we trained ourselves. You can equally pair `memory-manager/` with an untrained Qwen-7B, GPT-OSS-120B, or any other instruction-tuned LLM as the answer agent — see `tab:different-answer-agent` in the paper. The reverse never happens: `answer-agent/` is not used for memory operations, and swapping it out has no effect on how memory is maintained.
+**This model only manages memory — it does not answer questions.** A separate answer agent reads the memory store this model produces and generates answers; it can be any instruction-tuned LLM. Our own SFT+RL-trained answer agent is released separately at **[ahmedehabb/Memory-R2-answer-agent](https://huggingface.co/ahmedehabb/Memory-R2-answer-agent)**.
 
 ## Headline results (`tab:main`)
 
-`memory-manager/` is the constant in every row below; only the answer agent changes:
+This memory manager is held constant; only the paired answer agent changes:
 
-| Memory manager | Answer agent | F1 | BLEU-1 | LLM-judge (gpt-4o-mini) |
-| --- | --- | ---: | ---: | ---: |
-| memory-manager/ | `answer-agent/` (ours, SFT+RL) | **51.46** | **44.84** | 69.03 |
-| memory-manager/ | GPT-OSS-120B (untrained, external) | 49.29 | 43.64 | **86.08** |
+| Answer agent | F1 | BLEU-1 | LLM-judge (gpt-4o-mini) |
+| --- | ---: | ---: | ---: |
+| [ahmedehabb/Memory-R2-answer-agent](https://huggingface.co/ahmedehabb/Memory-R2-answer-agent) (ours, SFT+RL) | **51.46** | **44.84** | 69.03 |
+| GPT-OSS-120B (untrained, external) | 49.29 | 43.64 | **86.08** |
 
-## Usage
+See the paper's `tab:different-answer-agent` for more pairings (untrained Qwen-7B, etc.) — the memory manager is not tied to any one answer agent.
 
-These models are two halves of a two-agent pipeline (memory manager + answer agent) and are not intended as general-purpose chat models on their own. Full inference code and the memory-store protocol are in the [project repository](https://github.com/) (see the paper for the official release).
+## Usage
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-memory_manager = AutoModelForCausalLM.from_pretrained("ahmedehabb/Memory-R2", subfolder="memory-manager", torch_dtype="auto", device_map="auto")
-answer_agent = AutoModelForCausalLM.from_pretrained("ahmedehabb/Memory-R2", subfolder="answer-agent", torch_dtype="auto", device_map="auto")
+memory_manager = AutoModelForCausalLM.from_pretrained("ahmedehabb/Memory-R2", torch_dtype="auto", device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained("ahmedehabb/Memory-R2")
 ```
 
+Full inference code and the memory-store protocol are in the [project repository](https://github.com/ahmedehabb/Memory-R2) (see the paper for the official release).
+
 ## Training
 
 - Base model: `Qwen/Qwen2.5-7B-Instruct`
-- Memory manager: LoGo-GRPO (turn-level + token-level advantage), curriculum-trained 8-session → 16-session → 32-session; reward = per-session cumulative F1 against gold QA + a memory-compression penalty (λ=0.3)
-- Answer agent: SFT warm-start followed by an RL continuation (answer-F1 reward) against the memory manager's rollouts
+- Algorithm: LoGo-GRPO (turn-level + token-level advantage), curriculum-trained 8-session → 16-session → 32-session
+- Reward: per-session cumulative F1 against gold QA + a memory-compression penalty (λ=0.3)
 - Judge for reward/logging during training: GPT-OSS-120B
 
 ## Citation