iHow Memory

An open spec draft for durable, auditable memory across AI agents

中文 · Whitepaper · Protocol · Scenarios · Diagrams

AI agents are powerful inside one chat. Real work is longer than one chat.

iHow Memory defines a local-first memory and handoff reliability layer so agents, tools, and human operators can share durable project context across sessions, models, and handoffs without relying on hidden chat history.

v0.1 is an open spec draft seeking RFC-style review. It intentionally publishes specifications, scenarios, diagrams, and documentation only. No implementation code is included.

v0.1 first published by iHow1, 2026-05.

The Problem in One Picture

flowchart LR
  U["Human operator"] --> A["Agent A works"]
  A --> E["Events: decisions, files, feedback, blockers"]
  E --> W["Writeback: proposed durable memory"]
  W --> R["Review: approve, revise, reject"]
  R --> M["Project memory: scoped and auditable"]
  M --> C["Context package for the next task"]
  C --> B["Agent B continues without starting over"]
  M --> AU["Audit: why was memory used?"]

Loading

Why This Matters

Most AI work fails at the handoff boundary:

Failure	What happens
Revision amnesia	The same feedback is repeated again and again.
Tool-swap amnesia	Moving from one AI tool to another loses project state.
Handoff amnesia	A new person or agent must reread raw history before becoming useful.
Safety drift	Hard constraints become ordinary suggestions and get ignored.

iHow Memory treats memory as shared project infrastructure, not a private feature inside one agent.

Who Is This For?

• Teams using multiple AI coding or writing tools on long-running projects.
• AI product builders who need memory semantics beyond chat history.
• Operations, support, and consulting teams that rely on repeatable handoffs.
• Researchers and analysts who need durable project state across tools.
• Organizations that need local-first, auditable AI workflow memory.

What Makes It Different

• Local-first by default: project memory stays in the operator's chosen environment.
• Human-readable: durable state is inspectable and reviewable.
• Model-neutral: memory semantics are not tied to one LLM provider.
• Multi-agent native: handoff is a first-class reliability target.
• Auditable: memory has provenance, scope, review status, and lifecycle records.
• Conformance-oriented: quality is measured by behavior, not by storage technology.

Comparison With Other Approaches

Approach	Primary focus	iHow Memory difference
Chat history	Recall inside one product session	Defines durable project memory across sessions and tools.
Vector database / RAG	Retrieve similar text	Adds scope, review status, hard constraints, writeback, and audit.
Agent memory platforms	Personalization or app-level memory	Focuses on multi-agent handoff reliability and conformance.
Manual project docs	Human-maintained notes	Defines protocol semantics for agents to record, retrieve, and audit memory.

v0.1 Published Materials

Area	File
Non-technical overview	docs/overview-for-non-technical-readers.md
Diagrams	docs/diagrams.md
Whitepaper	whitepaper/whitepaper-public-v0.1.en.md
Chinese whitepaper	whitepaper/whitepaper-public-v0.1.zh-CN.md
Protocol draft	spec/protocol-draft-v0.1.md
Reliability scenarios	scenarios/reliability-scenarios-v0.1.md
Conformance direction	conformance/README.md
Release scope	docs/release-scope-v0.1.md
Security boundary	docs/security-boundary.md

Reliability Scenarios

The v0.1 scenario set defines five acceptance-style tests:

1. Cross-Tool Handoff / 跨工具接力
2. Feedback Pattern Capture / 反馈规律沉淀
3. Constraint Preservation / 禁忌约束执行
4. Human Team Handoff / 新人接手
5. Model Migration / 跨模型迁移

Each scenario includes Given, When, Then, failure modes, and acceptance criteria.

Protocol Draft

The protocol draft defines four core interfaces:

• events: workflow event ingestion
• context: bounded context package retrieval
• writeback: proposed durable memory and review
• audit: traceability and lifecycle control

It also defines isolation boundaries for tenant, customer, project, and user scopes.

Validation Direction

The public v0.1 repository publishes a reliability language first: reviewable scenarios, protocol semantics, isolation boundaries, and conformance direction. It does not publish benchmark results or executable conformance tooling. Future versions may consider executable checks after the spec boundary is stable.

What Is Not Included in v0.1

This repository intentionally does not include:

• tool integration code
• SDKs
• runtime services
• hosted service code
• deployment recipes
• private operations
• customer-specific materials
• generated benchmark data

The first public release is deliberately narrow: define the reliability language before publishing implementation details.

License

• Specification and scenario materials are licensed under CC BY 4.0. See LICENSE-SPEC.
• Whitepaper and documentation materials are licensed under CC BY 4.0. See LICENSE-DOCS.
• The iHow Memory name and marks are not licensed for unrestricted brand use. See TRADEMARK.md.

Future code releases, if any, may use a separate software license. No software license is granted by this v0.1 repository.

Get Involved

Help shape the spec draft:

• Open an issue for ambiguous terminology, missing failure modes, or security boundary concerns.
• Suggest additional reliability scenarios.
• Review the conformance direction and propose behavior-level checks.
• Share implementation feedback without publishing private code or customer materials.

Status

v0.1 draft for review.