A Lifecycle Memory Framework for Personal AI
Homepage • Overview • Architecture • Highlights • Results • Citation
- 2026-07 — Technical report posted on arXiv: Mi-Memory: A Lifecycle Memory Framework for Personal AI.
- 2026-07 — Repository currently hosts the project introduction page, report PDF, and overview figures.
Project homepage: https://darwin-agent.github.io/Mi-Memory/
This repository currently hosts the Mi-Memory technical report, project homepage, and visual assets.
Personal AI is moving beyond chat-only interaction toward continuous services across phones, cars, homes, wearables, cameras, and tools. In this setting, memory cannot remain a cache of prior conversations: it must preserve durable user state, connect answers to multimodal and device evidence, support correction and forgetting, bound policy evolution, and remain deployable across edge/cloud constraints.
Mi-Memory is a lifecycle memory framework organized around four roles:
- Structure — memory runtime, storage hierarchy, retrieval, filtering, and context assembly.
- Expansion — multimodal and cross-device evidence acquisition.
- Evolution — diagnostic iteration, governed strategy updates, and rollback.
- Deployment — substrate-independent memory variants for constrained environments.
Mi-Memory links the lifecycle through a shared audit contract: typed evidence payloads preserve source identity and provenance, diagnostic traces localize evidence loss, strategy artifacts make memory-policy changes explicit, and gate/rollback records bound accepted evolution.
| 🧱 Composable memory structure | Multi-granularity storage with stage-level diagnostic traces. |
| 🌐 Multi-source evidence | Dialogue, multimodal perception, cross-device events, and causal fusion. |
| 🔄 Governed evolution | D²ACCI and E²MEND governed strategy evolution. |
| 📦 Deployment flexibility | Full-stack cloud reference and repository-native lightweight memory variant. |
| 📊 Comprehensive evaluation | Structure benchmarks plus module-level, transfer-feasibility, and design-level evidence. |
The results below summarize evidence anchors from the technical report. They are reported role by role rather than as a single leaderboard, and evidence maturity differs across tracks. See the arXiv report for the authoritative numbers, protocols, evidence boundaries, and detailed analysis.
| Benchmark | Track | Metric | Score |
|---|---|---|---|
| LoCoMo | Structure | Judge Accuracy | 93.59% |
| PersonaMem-V2 | Structure | Preference Accuracy | 57.24% |
| LongMemEval | Structure | Judge Accuracy | 87.47% |
| Mem-Gallery | Expansion | Judge Accuracy (3-vote) | 89.15% |
| MemFuseBench | Expansion | Internal Fusion Score | 35.2% |
| D²ACCI / E²MEND LoCoMo | Evolution | Offline Run Accuracy | 94.74% |
| LiteMem / LoCoMo-aligned | Deployment | Transfer Score / Retention | 90.81% / 90.0% |
Mi-Memory is an initial step toward auditable personal-AI memory infrastructure. Future work will focus on stronger causal attribution, propagation-complete forgetting, federated cross-device memory, scalable diagnostic evolution, and standardized memory contracts.
.
├── index.html # GitHub Pages project homepage
├── paper.pdf # Released technical report
├── figure/ # README-facing visual assets
├── LICENSE
└── README.md
This repository is an introduction page for the Mi-Memory technical report.
If you find this work useful, please cite:
@techreport{mimemory2026,
title = {Mi-Memory: A Lifecycle Memory Framework for Personal AI},
author = {Darwin Agent Team},
institution = {Xiaomi},
year = {2026},
url = {https://arxiv.org/abs/2607.18975}
}This project is licensed under the MIT License.