Daily → Weekly Compaction Prompt

Spec reference: §5.2 (Daily → Weekly), §5.3 (Typed Memory Extraction)
Used by: Compactor.compactDaily(week)
Model call: MemoryModel.complete(prompt, { systemPrompt, temperature: 0.2, maxTokens: 4096 })

System Prompt 

You are a memory compaction engine. You compress daily agent logs into a structured weekly summary. You also extract durable knowledge into typed memory candidates.

Rules:
- Output ONLY the two sections described below. No preamble, no commentary.
- Use the exact markdown structure shown. Do not invent additional sections.
- Compression target: the WEEKLY SUMMARY section should be roughly 30% of the combined input token count.
- Every claim in your output must trace to a specific daily entry. Do not hallucinate or infer beyond what is written.

User Prompt Template 

<DAILY_LOGS>
{daily_logs}

<INSTRUCTIONS>
Compact the daily logs above into two sections.

SECTION 1 — WEEKLY SUMMARY

Write a structured markdown summary for the week. Use this exact format:

---
type: weekly
---

## {week_label}

### Key Decisions
- (list decisions made, with date and rationale)

### Outcomes
- (list what was accomplished, shipped, merged, or resolved)

### Blockers & Open Items
- (list unresolved issues, blockers, or items carried forward)

### Context
- (any other detail worth preserving at the week level)

Rules for this section:
- DROP routine entries (status checks, trivial updates, repeated daily patterns)
- KEEP decisions, outcomes, blockers, surprises, and anything a future reader would need
- Each bullet should be self-contained — readable without the original daily log
- Target ~30% of input length

SECTION 2 — TYPED MEMORY CANDIDATES

For each entry in the daily logs that represents durable knowledge, emit a fenced block:

```typed_memory
---
name: {short_title}
description: {one-line description for search indexing}
type: {user | feedback | project | reference}
---

{content — for feedback/project types, include **Why:** and **How to apply:** lines}

Classification rules:

  • user: facts about a person’s role, preferences, expertise, or working style
  • feedback: guidance on how to approach work — corrections OR validated approaches. Include why.
  • project: decisions, goals, timelines, or context about ongoing work. Convert relative dates to absolute.
  • reference: pointers to external resources (URLs, tools, dashboards, channels)

If NO entries qualify as typed memories, output:

NONE

Do not extract memories that are:

  • Derivable from code or git history
  • Ephemeral task state (in-progress work that will change tomorrow)
  • Already documented elsewhere ```

Parsing the Output

The compactor parses the response into two parts:

  1. Weekly summary — everything outside typed_memory fenced blocks → written to memory/weekly/{week}.md
  2. Typed memory candidates — each typed_memory block is parsed for frontmatter and written to memory/{type}_{topic}.md

The typed_memory fence language tag makes extraction unambiguous — simple regex: typed_memory\n([\s\S]*?)

Design Rationale

Principle Application
Positional attention System prompt sets role + hard rules at the top. Daily logs (data) come next. Instructions at the bottom — closest to where the model generates output.
Labels adjacent to values The output template puts field names (name:, type:) directly next to their values in the typed memory blocks.
Force discrete outputs Type classification is constrained to 4 options. The keep/drop decision is implicit in the output structure — if it’s not in the summary, it was dropped.
Compress before reasoning The model doesn’t analyze the logs first and then summarize. It reads compressed daily logs and directly produces the compressed output.
Scope retrieved context Only the relevant week’s daily logs are injected. No extra context.
Open-only section tags <DAILY_LOGS> and <INSTRUCTIONS> use open-only tags. The next tag implicitly closes the previous section.