Temporal Summary

Temporal Summary（时间摘要）

Objective, indiscriminate digests of conversation activity over time buckets. Related: memory.md, sleep.md, diary.md, compression.md, entity-model.md.

命名

时间摘要 (temporal_summary) — digests of conversation content, bucketed by day / month / year. Not a memory taxonomy entry (semantic / limbic / narrative), not diary prose, not the runtime compression summary segment.

vs Diary

Axis	Diary	Temporal summary
Stance	Subjective	Objective
Selection	Editorial	Indiscriminate coverage of active sessions
Storage	diary_entry + blocks	Global → entity; today per-conversation → conversations.temporal_day
Primary reader	/diary UI	System prompt (yesterday+) / timeline inject (today peers)

LLM compression still applies (char caps). Objective ≠ exhaustive log replay: cover events without an editorial “worth remembering” filter, but write event-level digests (themes + outcomes). Omit internal IDs, step-by-step tool/config actions, and per-notification timestamps unless essential to understand what happened.

Storage layers

Layer	Where	Citable ([[anima:id]])
Global day / month / year	entities primary_component=temporal_summary	Yes
Conversation day chunks (today)	conversations.temporal_day JSONB	No (operational)

Conversation JSONB

{
  cst_date: string,
  chunks: Array<{
    at: string,
    bucket: string, // CST half-hour bucket start, e.g. 2026-07-18T06:00+08:00
    summary: string,
    watermark_message_id?: string,
    watermark_at?: string,
  }>
}

Chunks are append-only within a CST day. Tick only pushes new chunks when there are messages after the watermark.

Global entity body

{ window: "day" | "month" | "year", period_start: string } // CST YYYY-MM-DD

Unique on (window, period_start) for global rows (expression unique index).

Generation

Step	Trigger	Output
Conversation chunk	in-process Bun.cron builtin-temporal-summary-tick */30	Append chunk if new messages
Peer rollup	Same tick / on assemble for closed buckets	One merged peer digest per viewer source-set → Redis cache
Global day	Sleep step temporal-summary-day (after light-sleep)	Overwrite global day entity for sleep day
Month / year	Sleep step temporal-summary-cascade (after deep-sleep)	Month on month-end; year on Dec 31

Identity context (self + resident) must ride with LLM summarization calls.

Redis peer rollup key

{prefix}:peer_roll:{cst_date}:{bucket}:{sources_fp}

• Default prefix: anima:temporal (config memory.temporal_summary.redis_key_prefix)
• sources_fp: short hash of canonical sorted (conversation_id, at, summary) excluding the viewer
• Value: { summary, sources_fp, created_at }
• TTL: ~36h (discardable cache via Redis cache layer + in-process fallback)
• Miss: LLM merge (or concatenate truncate if LLM unavailable); Hit: reuse

Same source set shares one key across viewers that exclude the same peers.

Injection（LLM prefix / KV cache）

Content	Where	When it changes
Yesterday … earlier (rolling)	System prompt section temporal-summary	CST 02:00 rebuild only
Today other sessions	Not system	Closed half-hour buckets only

Today: timeline insert (one block per closed bucket)

When assembling messages for viewer V:

1. For each closed CST half-hour bucket today, collect other sessions’ chunks in that bucket.
2. Resolve peer rollup via Redis key above → one summary string.
3. Insert a runtime-only assistant message (name: temporal_summary_peers) at bucket_end on V’s timeline (by message timestamps).
4. Never rewrite older buckets; new activity only opens newer buckets → stable prefix for LLM KV cache across parallel sessions.

Forbidden: putting today peers in system prompt; per-turn tail replace of a stable today block; injecting N raw peer chunks side-by-side for one bucket.

配置

memory.temporal_summary in Habitat runtime config (habitat_runtime_config): enabled, char caps, redis_key_prefix.

Sleep relationship

Global day overwrite is a side product of the sleep cycle, not a replacement for light-sleep semantic / limbic / autobiography stages.