Forgetting engine for AI agents — memory TTL, decay, eviction, audit. The opposite of memory programming.
Server defines 7 tools with complete input schemas and descriptions in Chinese. Naming follows verb_noun pattern (amnesia_remember, amnesia_forget, amnesia_recall). Schemas are well-structured with enums, required fields, and type definitions. However, descriptions lack English translations (barrier to LLM understanding in English-first systems), and output schemas are minimally documented. Error handling guidance is absent, tools do not indicate what failures mean or how to recover. Parameter descriptions are present but terse (avg ~40 chars). No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels (WRITE, DESTRUCTIVE, READ_ONLY). Missing pagination guidance for audit/recall results.
审计日志:最近 N 条遗忘/恢复事件(id、kind、reason、at、contentHash;restorable 模式含 content)。
拔下记忆:按 id / 内容关键词 / 标签 / 全部 遗忘。遗忘物理删除内容并写入审计(只留哈希;restorable 模式保留副本可恢复)。
干跑预览:当前预算(maxInjectedTokens)下下一步将注入哪些记忆、占多少 token、跳过多少死记忆(不 touch、不生效)。token 成本对账用。
想起:按关键词召回记忆并触摸(滑动续期)。被召回的每条记忆衰减时钟重置——只要持续有用就不过期(use it or lose it)。
记住一条带保质期的记忆(插上记忆总线)。默认按衰减曲线在 TTL 后过期;pin=true 则永久保留,直到显式遗忘。
从审计恢复一条被遗忘的记忆(restorable 模式)。返回新 id、新 TTL;恢复动作本身记入审计。隐私模式(restorable=false)下内容已物理删除,无法恢复,只能重新 remember。
记忆健康度读数:活跃/陈旧/已遗忘数量、token 总占用、下一次过期时间、上次注入账本(count/tokens/at)。
Descriptions are in Chinese only; no English translations. LLMs trained primarily on English will struggle to understand tool purpose and selection criteria.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite explicit risk labels in metadata (WRITE, DESTRUCTIVE, READ_ONLY). LLMs cannot infer safety properties.
Output schemas minimally documented. amnesia_remember render() shows example output but schema is {type: 'object', additionalProperties: true}, no field definitions. Agents cannot plan downstream calls.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
amnesia_forget accepts 'all: true' with 'confirm: true' guard but no error message guides recovery if confirm is omitted. Error handling lacks actionable guidance.
amnesia_audit and amnesia_recall return unbounded lists (limit defaults to 50 but no max enforced). Large result sets risk context window exhaustion.