trove-remember

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

trove-remember

"Save the gold — the graph decides whether it's new or a change of mind."

"留存有价值的信息——由图谱判断其为新内容还是观念变更。"

When to use

使用场景

Mid-session when a decision, root cause, gotcha, or durable fact crystallises (do not wait for wrap-up).
After a debugging or design session that produced new understanding.
The user stated a fact, constraint, or preference to remember.
A known fact changed: "PR merged — mark it done", "that port is 5433, not 5432", "X replaced Y".
End-of-session pass: 3–8 high-value atoms with links — not one mega session-summary node.

When not to use:

A specific external source to index →
```
trove-ingest
```
first, then remember distilled atoms citing text units.
A question to answer →
```
trove-recall
```
.
Trivial Q&A / ephemeral debugging chatter.

会话进行中：当决策、根本原因、注意事项或持久化事实明确时（不要等到会话结束）。
调试或设计会话产生新认知之后。
用户提出需要记住的事实、约束条件或偏好时。
已知事实发生变化时：「PR已合并——标记为完成」、「端口是5433，不是5432」、「X替代了Y」。
会话结束阶段：留存3–8个带链接的高价值信息单元——不要创建一个巨型会话总结节点。

Process

禁用场景

Step 1 — scan and propose

—

Collect candidates: decisions ("we'll do X because Y"), facts (values, IDs, paths, ports stated as authoritative), gotchas ("this fails if…"), preferences ("always…", "never…"). Present a short numbered checklist; the user picks. Be conservative — 5 high-quality atoms beat 30 noisy ones. For a single explicit correction ("port is 5433"), skip the proposal and just do it.

需要索引特定外部资源 → 先使用
```
trove-ingest
```
，再引用文本单元留存提炼后的信息单元。
需要回答问题 → 使用
```
trove-recall
```
。
琐碎问答/临时调试对话。

Step 2 — remember

操作流程

—

步骤1 — 扫描并提议

One call per atom:

remember { title, type, summary, content?, evidence, links }

```
remember
```
searches for an exact title/slug match itself: match → new revision of that node; no match → new node. No baseRevisionId to manage.
Check the response.
```
action
```
tells you what happened;
```
similar
```
lists near-matches it did NOT merge into. If one of those is the node you meant, re-call with
```
slug: <that-slug>
```
to force the revision.
Evidence: cite
```
sourceId
```
/
```
textUnitId
```
when the fact came from a document; otherwise say "agent inference from session <date>" in the summary.
Links: connect to the relevant project/domain nodes via
```
links: [{toSlug, predicate}]
```
.

收集候选内容：决策（「我们将做X，因为Y」）、事实（权威陈述的值、ID、路径、端口）、注意事项（「如果…会失败」）、偏好（「始终…」、「绝不…」）。呈现简短的编号清单供用户选择。宁缺毋滥——5个高质量信息单元胜过30个冗余内容。对于明确的单一修正（「端口是5433」），可跳过提议环节直接执行。

Step 3 — supersede relationships (only when a belief between nodes changed)

步骤2 — 留存信息

connect { fromNodeId, toNodeId, predicate, supersedesEdgeId }

— atomically creates the new edge and expires the old one.

Belief retired with no replacement:
```
forget { edgeIds: [...] }
```
, or
```
forget { query, dryRun: true }
```
to preview first.
Find edge ids via
```
read
```
(connections) or
```
neighborhood
```
.
Never express change by deleting:
```
asOf
```
time-travel depends on the trail.

每个信息单元调用一次：

remember { title, type, summary, content?, evidence, links }

```
remember
```
会自行搜索精确的标题/别名匹配：匹配成功→更新该节点的新版本；无匹配→创建新节点。无需管理baseRevisionId。
检查响应结果。
```
action
```
字段会告知操作结果；
```
similar
```
字段列出未合并的近似匹配项。如果其中某个是目标节点，可使用
```
slug: <对应别名>
```
重新调用以强制修订。
证据：若事实来自文档，需引用
```
sourceId
```
/
```
textUnitId
```
；否则在摘要中注明「Agent从<日期>的会话中推导得出」。
链接：通过
```
links: [{toSlug, predicate}]
```
连接到相关项目/领域节点。

Step 4 — propagate

步骤3 — 替换关联关系（仅当节点间的认知发生变化时）

neighborhood { nodeId, depth: 1 }

on changed nodes — adjacent summaries restating the old fact get their own

remember

(slug-targeted) or get flagged to the user.

connect { fromNodeId, toNodeId, predicate, supersedesEdgeId }

— 原子化创建新边并将旧边标记为过期。

无替代方案的认知退役：使用
```
forget { edgeIds: [...] }
```
，或先执行
```
forget { query, dryRun: true }
```
预览结果。
通过
```
read
```
（连接）或
```
neighborhood
```
查询获取边ID。
绝不要通过删除表达变更：
```
asOf
```
时间回溯依赖完整的变更轨迹。

Step 5 — mirror

步骤4 — 传播更新

If the fact belongs on a human-readable vault page, run the matching

/scribe-*

flow. Note that vault import is one-way (vault → graph); graph edits don't write back to the vault automatically.

对变更节点执行

neighborhood { nodeId, depth: 1 }

——重述旧事实的相邻摘要需进行针对性的

remember

（指定别名）或标记给用户。

Step 6 — confirm

步骤5 — 同步镜像

One sentence: how many atoms landed (created vs revised), with slugs.

若该事实需要同步到人类可读的Vault页面，执行对应的

/scribe-*

流程。注意Vault导入是单向的（Vault → 图谱）；图谱编辑不会自动写回Vault。

Anti-patterns

步骤6 — 确认结果

Don't save without proposing first when harvesting a session — the user picks. (Direct corrections are exempt.)
Don't capture speculation as fact; decisions are things the user agreed to.
Don't ignore
```
similar
```
— a near-match you should have merged into is graph rot.
Don't rewrite a whole node when a summary tweak suffices; pass only the fields that changed.
Don't delete edges. Supersede (
```
connect
```
) or retire (
```
forget
```
).

用一句话说明：成功留存的信息单元数量（创建vs修订），并附上别名。

—

反模式

—

不要在会话梳理时跳过提议环节直接保存——需由用户选择。（直接修正除外。）
不要将推测内容作为事实保存；决策必须是用户认可的内容。
不要忽略
```
similar
```
字段——本该合并的近似匹配项会导致图谱冗余。
不要在只需调整摘要时重写整个节点；仅传递发生变化的字段。
不要删除边。应使用替换（
```
connect
```
）或退役（
```
forget
```
）操作。