Memory & Persona Runtime for AI Agents
Eidema originates in Amadeus and is designed for reuse by other host applications. V1 focuses on memory; core persona settings remain with the host.
This public branch contains design documents and selected progress reports. Local development has implemented the M0 source contract, M1 explicit-memory lifecycle and recall foundations, and an exact source-quote locator. This update does not publish the runtime code, provide an installable release, or deploy a new policy.
The technology direction remains Python + SQLite. Original sources, explicit memories, automatically proposed memories and derived views share source, scope, version and invalidation semantics. FTS and multilingual vectors are retrieval indexes, not separate authorities.
- Current progress and implementation boundaries (中文)
- Important decisions and their limits (中文)
- Selected measurements, denominators and caveats (中文)
- Next implementation steps (中文)
- Reading Anthropic's “The components of agent memory” (中文)
The explicit lifecycle is a local engineering foundation. P150-F is a conditionally retained research baseline, not an approved ordinary-chat default. Automatic formation, warm-path preparation and cold-path reflection remain incomplete. The prepared 64-request formation-contract comparison was deferred by the user and has made zero formal requests.
Only compact reports, aggregate metrics and archive hashes are published here. The full local evidence archives are not included; this is not a complete reproduction package.
- Remember, inspect, correct and forget across sessions, with identifiable sources and access scope.
- Deleting a chat and forgetting an explicitly saved memory are distinct operations.
- Preserve original evidence; extracted claims and summaries do not become independent proof of themselves.
- Models interpret meaning. The host owns identity, clocks, permissions, execution authority and authoritative application state.
- Memory does not gain system-level authority. Persona definitions, TTS settings, permissions and Work/Schedule execution remain host-owned.
- Tags and entity/event views may aid retrieval without prescribing a fixed classification route.
- Background work must be bounded, recoverable and invalidated when its sources change. Forgetting and permission changes must not wait for cold maintenance.
- Relationship growth, learned interaction habits and automatic persona evolution remain future work.
Original-source retrieval and explicit memory remain the comparison baseline for formation and consolidation. Evaluation includes useful recall, subject attribution, unwanted intrusion, persona effects, and separately measured visible and spoken response latency.
These pages preserve the original design discussions. Their proposed fields and budgets are not a claim that every capability is implemented; use the dated progress reports above for current status.
- V1 memory responsibilities and boundaries
- Persistent-agent scheduling boundaries
- V1 data and runtime proposal
- V1 data model
- V1 retrieval and context proposal
- Benchmark references and coverage gaps
- Runtime paths and persona interaction notes
A project license has not yet been selected. Reuse is a design goal, not an assertion that an open-source license has already been granted.
Eidema面向持久agent的跨会话记忆与后续人格连续性，先服务本地Amadeus，同时保留其他宿主接入能力。第一版先把记忆基础做扎实，人格成长后置。
来源身份与修订、显式记忆生命周期、检索及本轮来源引文定位已有本地实现；温冷路径和自动形成尚未完整落地。本次远程更新只发布精选文档，没有推送运行时代码、迁移真实记忆库或打开普通聊天自动注入。
当前重点是让原话、候选和派生视图具有共同的依据与失效规则，再逐步推进形成、局部整理和低频反思。既看需要时能否找对、用对，也看无关时是否安静；不能把检索命中当成回答正确。
详见本阶段快照。其中明确区分已实现、已测候选、未通过的结果与尚未运行的实验。