You can deploy this model and start earning money today!

2009YU/
TMCRA-Agent-Memory

Name: TMCRA-Agent-Memory
Availability: InStock

information-retrieval

longmemeval

apache-2.0

Views

No views yet

TMCRA — Agent Memory Engine

TMCRA logo

简体中文 · GitHub · Release v2026.07.18

TMCRA is a memory engine for long-running agents. It converts multi-turn and multi-session conversations into scope-isolated, source-traceable memory, then recalls compact evidence after the user's next question arrives.

This Hugging Face repository contains the Apache-2.0 release of the TMCRA algorithm, its LongMemEval reproduction pipeline, and the inference-only graph scoring weights used by the released benchmark chain. Hosted APIs, accounts, billing, and production control-plane services are not included.

LongMemEval S500

TMCRA achieved 411 / 500 = 82.2%.

Task category	Correct / Total	Accuracy
Knowledge Update	71 / 78	91.0%
Multi-session	90 / 133	67.7%
Single-session Assistant	55 / 56	98.2%
Single-session Preference	27 / 30	90.0%
Single-session User	67 / 70	95.7%
Temporal Reasoning	101 / 133	75.9%
Overall	411 / 500	82.2%

The machine-readable scorecard is results/latest_benchmark.json. The maintained reproduction entrypoint is benchmarks/longmemeval/. The benchmark answer layer is fixed to GPT-5.4; fresh runs that call external model services can vary slightly.

Architecture

TMCRA separates memory construction from downstream answer generation.

text

Conversation history
        │
        ▼
      Writer ─────────────── subject / entity attribution
        │
        ├── Source layer ─── immutable, traceable evidence
        ├── Fast layer ───── atomic current-state memory
        └── Slow layer ───── durable semantic graph
                 │
                 └────────── scope-bound indexes

New question
        │
        ▼
Recall Planner → layered retrieval → node/path scoring → reranker
        │
        ▼
Source-bound evidence packet → downstream agent

The TMCRA core ends at the evidence packet. The answer model and benchmark judge are harness components, not part of the online memory engine. User statements and assistant progress remain separately attributed, while both can be recalled when relevant.

Released weights

All three checkpoints contain only their public inference contracts and model state. Training state, machine paths, and private run metadata are excluded. Load them with torch.load(..., weights_only=True) through the repository runtime.

File	Purpose	Size	SHA-256
`node_scorer.pt`	Graph node scoring	207,356,243 bytes	`d2318aafb07f9a15a1d95e6eac1b3e09afa07c570dd9d15d3f75544c9950f201`
`path_scorer.pt`	Graph path scoring	207,356,243 bytes	`ecdc2bca51b7646a8a56c7db9cea4aeec581e64c3b61e6e06f7e9fa2a93b3bf5`
`tmcra_v3_reranker.pt`	Learned retrieval reranking	4,232,213 bytes	`09a285b484ca857b24b53ad19d5998302f16d224ca9e4fd73a4eb5a52f022942`

The files and their machine-readable contract are under models/tmcra_v4_longmemeval_s500_20260715/.

Download

Install the current Hugging Face CLI and download only the released inference package:

bash

python -m pip install -U huggingface_hub

hf download 2009YU/TMCRA-Agent-Memory \
  --include "models/tmcra_v4_longmemeval_s500_20260715/*" \
  --local-dir ./TMCRA-Agent-Memory

For source, benchmark setup, environment examples, asset checksums, and stage validation, use the GitHub release or benchmarks/longmemeval/README.md.

Version history

Current release: v2026.07.18
Preserved pre-release snapshot: legacy-2026.05.25

Developers

Yu Haoxin — creator, lead developer, and TMCRA algorithm engineering.
OpenAI Codex — development and reproducibility engineering assistant.

See AUTHORS.md and CITATION.cff for attribution and citation details.

License

TMCRA is released under the Apache License 2.0. Third-party datasets, models, and components retain their own licenses.