MemWeaver — Memory Profiler

Your Agent reads your memory every day, but does it truly understand you?

MemWeaver digs into your memory files — long-term memory (MEMORY.md) and daily logs — to uncover preferences, behavioral patterns, and hidden traits you might not even be aware of, then confirms findings through an interactive questionnaire and outputs a structured user profile (YAML).

What Makes This Different

Core value: MemWeaver finds the gap between what you say you prefer and what you actually do — then asks you about it.

Overview

• - Input: MEMORY.md (long-term memory) + recent daily logs (memory/*.md)
• Process: LLM deep analysis → batch interactive questionnaire
• Output: Structured user profile at INLINECODE3

Workflow

Step 1: Collect Memory

CODEBLOCK0

The script reads long-term memory and recent logs, outputs JSON to stdout. The Agent parses content.long_term and content.daily_logs fields from the JSON.

Note: If estimated_tokens exceeds 8000, consider reducing the --days parameter.

Step 2: LLM Deep Analysis

The Agent analyzes collected memory in 3 sub-tasks:

2.1 Basic Profile Extraction

Extract factual information from memory across these dimensions:

Tag each field with a confidence score.

2.2 Hidden Pattern Mining

This is MemWeaver's most valuable part. The Agent specifically analyzes these 6 types of hidden patterns:

1. 1. Decision patterns: What does the user lean toward when facing multiple options? (Analysis-driven vs intuition? Fast vs slow decisions?)
2. Time & energy allocation: Does actual energy distribution (from log frequency) match user's self-description?
3. Overlooked interests: Topics that appear repeatedly but the user hasn't formally tracked
4. Statement vs behavior contradictions: Are stated preferences inconsistent with actual actions?
5. Emotion/energy triggers: What scenarios make the user especially productive or resistant?
6. Unlabeled skills: Abilities the user demonstrates but hasn't self-recognized

Each finding needs evidence (citing specific memory content) and reasoning logic.

2.3 Project Importance Re-evaluation

List every project and idea recorded in MEMORY.md, provide reassessment:

• - Current status (active / paused / archived / concept)
• Suggested importance (high / medium / low / shelved)
• Assessment reasoning (frequency in logs, recent activity, user investment)
• Questions to confirm with user (if uncertain)

Step 3: Interactive Confirmation (Batch Questionnaire)

Interact with the user in batch mode, similar to a personality test. Each question is based on the user's actual memory content, not just showing analysis conclusions.

Core design principles:

• - Push 5 questions per batch, user answers at once (e.g., "1A 2C 3B 4D 5A"), Agent provides feedback then pushes next batch
• Questions reference user's real memories as context
• Provide options (A/B/C/D or open-ended), user can choose or free-form
• Type B (hidden insight) questions should be ≥50% of total — this is MemWeaver's core value
• Three question types interleaved, not strictly separated by rounds

Question Design Rules:

The Agent designs 10-15 questions based on Step 2 analysis. Three types:

Type A: Scenario Recall (validate profile facts, ≤25%)

Reconstruct a real scene from memory, let user choose the best description.

CODEBLOCK1

Type B: Hidden Insight (core value, ≥50%)

This is MemWeaver's most important question type. The Agent uses specific evidence from memory to point out contradictions or blind spots between user's "self-perception" and "actual behavior".

Methodology:

1. 1. Find user's explicit statement (e.g., "I prefer X")
2. Find contradictory behavioral records (e.g., logs show consistently doing Y)
3. Present the contradiction to user, guide explanation via options
4. Options should include: acknowledge contradiction, deny, offer new explanation, other

CODEBLOCK2

Hidden insight mining directions (look for these 6 types of clues in memory):

1. 1. Statement vs behavior contradictions: Stated preferences inconsistent with actual actions
2. Time allocation truth: Log frequency/length reveals real energy distribution vs stated priorities
3. Silence signals: Topics in MEMORY that disappear from logs → possible priority drift
4. Energy fingerprint: Length differences across log types → reveals energy sources
5. Choice patterns: Consistent tendencies when user faces decisions
6. Unlabeled skills: Abilities demonstrated but not self-recognized

Type C: Priority Trade-off (re-evaluate project importance, ≤25%)

Create resource-constraint scenarios, force user to choose between projects, revealing true priorities.

CODEBLOCK3

Question count and batching:

• - Total 10-15 questions, pushed in 2-3 batches (~5 per batch)
• Batch 1 (5 questions): 1-2 warm-up Type A + 3 Type B (hidden insights)
• Batch 2 (5 questions): 3-4 Type B + 1-2 Type C (priority trade-offs)
• Optional Batch 3 (2-3 questions): follow-up questions based on unexpected findings from previous batches
• After each batch, Agent waits for user to answer all at once

Answer processing:

• - After user submits a batch (e.g., "1C 2B 3A 4B 5C"), Agent processes collectively
• Give 1 brief feedback per question, noting profile inference
• If an answer reveals new insight leads, add follow-up questions in next batch
• All answers recorded internally as profile evidence, aggregated into Step 4

Completion:

• - After the last batch, Agent outputs a brief profile summary (like a personality test result page)
• Then proceeds to Step 4

Step 4: Generate and Save Profile

Generate the confirmed profile as YAML and save via script:

1. 1. Agent generates complete YAML profile (see "Profile Template" below)
2. Save via script:

CODEBLOCK4

Or via stdin:

CODEBLOCK5

The script automatically backs up old profiles and saves to output/profile_YYYYMMDD.yaml.

Profile Template

CODEBLOCK6

Output

• - User profile: INLINECODE10
• Analysis cache: {baseDir}/cache/analysis_cache.json (future version)

Requirements

No external dependencies. Python 3.8+ standard library only.