Structure
Memory runtime, storage hierarchy, retrieval, filtering, and context assembly for multi-granularity state.
Technical Report · July 2026
Mi-Memory
A Lifecycle Memory Framework for Personal AI that turns memory from a conversation cache into auditable, multi-source, governable user-state infrastructure.
Memory for always-on personal AI.
Personal AI is moving beyond chat-only interaction toward continuous services across phones, cars, homes, wearables, cameras, and tools. In this setting, memory must preserve durable user state, connect answers to multimodal evidence, support correction and forgetting, bound policy evolution, and remain deployable across edge/cloud constraints.
Expansion
Multimodal and cross-device evidence acquisition with source identity, provenance, and fusion contracts.
Evolution
Diagnostic iteration, governed strategy updates, feature gates, rollback records, and auditable improvement.
Deployment
Substrate-independent memory variants for cloud, edge, and lightweight local environments.
Controlled referencePublic benchmark with a reproduced baseline under a shared harness.
Module-levelStandalone component on a public benchmark with its own harness.
Preliminary/internalInternal benchmark evidence with external validity still pending.
Transfer feasibilityCapability mapped to a lighter substrate and task style.
A shared audit contract across the lifecycle.
Mi-Memory links structure, expansion, evolution, and deployment through typed evidence payloads, diagnostic traces, strategy artifacts, and gate/rollback records. The same artifact families support research diagnostics and product-facing memory governance.
Lifecycle overview: Structure, Expansion, Evolution, and Deployment are connected by auditable memory artifacts.
Designed for correction, provenance, and transfer.
The framework decomposes assistant memory into testable roles: typed state makes failures localizable, source-aware expansion preserves evidence boundaries, governed evolution turns failures into rollback-capable hypotheses, and deployment tests whether audit obligations survive substrate transfer.
Human-editable memory
Correction, forgetting, and policy updates are treated as first-class lifecycle events rather than ad hoc prompt patches.
Evidence-aware answers
Retrieval and context assembly preserve where memory came from, why it was selected, and how it can be inspected.
MemStackStructures durable state with layered L0/L1/L2 memory, hybrid retrieval, bounded context assembly, and adapter contracts.
MemSenseTurns visual and multimodal observations into source-aware evidence rather than disposable attachments.
MemFuseLinks related observations across cars, homes, phones, wearables, and tools into coherent episodes.
LiteMemTransfers selected contract elements to Markdown/Git artifacts and file-tool retrieval for lightweight settings.
The report grounds the lifecycle with a human-car-home training handoff: a parent, a car route, a phone calendar, and a home camera jointly determine whether the assistant should remind the family about a basketball bag before entering the expressway.
ObserveCalendar, vehicle route state, and home camera evidence arrive with source identity and timestamps.
FuseThe system links training time, travel feasibility, and the observed bag location into one episode.
AnswerThe assistant retrieves the fused episode and answers with inspectable supporting evidence.
ImproveUser corrections become gated memory-policy updates rather than hidden prompt or retrieval edits.
Evaluation highlights.
The report evaluates Mi-Memory across structure benchmarks, expansion modules, governed offline evolution, and deployment transfer. These numbers are evidence anchors rather than a single leaderboard; the PDF is authoritative for protocols and boundaries.
D²ACCI/E²MEND improves a LoCoMo offline run from 75.58% to 94.74%.
LiteMem is a deployment path, not a benchmark, and reports 90.81% transfer plus 90.0% retention.
Mi-Memory is an initial step toward personal-AI memory that can be measured, governed, and improved across its full lifecycle.
Causal attribution and forgetting
Future work targets stronger answer-to-evidence tracing and propagation-complete forgetting across derived summaries, profiles, and caches.
Federated and standardized contracts
Cross-device memory composition, scalable diagnostic evolution, and shared memory interfaces remain key research directions.
Citation.
If this report is useful for your research or product work, please cite Mi-Memory.
@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}
}